# EAGLE4 catalog (4801 entries, generated 2026-10-26 04:09)
Machine form: catalog.json. Query: `python3 kg_query.py <words>`.


## site_page (139)

- page:eagle4-white-paper | 2026-10-23 | current | EAGLE4 white paper
  https://hardware.research.yshicarrick.com/p/eagle4-white-paper/
  Best rounds, Llama-3.2-3B humaneval 20x512: H200 + four V80 drafter 846.20 tok/s vs H200 + its own EAGLE drafter 741.27 on the same w2 d15 tree (+14%); 851.50 vs 301.41 for the target alone; +20% on d
- page:t2-tested-progress-20261022b | 2026-10-22 | current | Track 2: latest tested progress on talon (2026-10-22 evening)
  https://hardware.research.yshicarrick.com/p/t2-tested-progress-20261022b/
  Two talon windows ran on 10-22. Window 1 tested 0xE4B7130E, the v8 image plus a ring link fix. The ring BIST passed 16/16, Llama 1B/3B/8B TP8 were bit-exact with 0 all-reduce retries, and Qwen3.8 matc
- page:t2-qwen27b-eagle-talon2-20261023 | 2026-10-23 | current | Qwen3.8-27B + EAGLE3 drafter on talon2: GPU vs 4-card FPGA drafter (2026-10-23)
  https://hardware.research.yshicarrick.com/p/t2-qwen27b-eagle-talon2-20261023/
  Same host (talon2, one NVIDIA H200 NVL, SGLang 0.5.16 eaglefit), same target (Qwen/Qwen3.8-27B bf16), same drafter (BitNet/SubLN EAGLE3 step 21491 from lab-host). Three ways to decode batch-1: target alon
- page:track1-qwen27b-4c1-20261023 | 2026-10-23 | current | Track 1: Qwen3.8-27B target with its drafter on the 4-card V80 FPGA drafter (4C1)
  https://hardware.research.yshicarrick.com/p/track1-qwen27b-4c1-20261023/
  <b>v2 (2026-10-23): the headline comparison is now same-day, same-H200 on talon2, humaneval only.</b> v1 compared lab-host FPGA runs with lab-host GPU runs from 2026-10-20 that were thermally throttled and ab
- page:eagle4-home | 2026-10-23 | current | EAGLE4 home: three tracks [0xE4B711C8,0xE4B711C9,0xE4B711CA,0xE4B711CD...]
  https://hardware.research.yshicarrick.com/p/eagle4-home/
  Start here: three track columns with headline numbers and daily progress blocks; tabs for the architecture book, model weights, knowledge base, research and operations. || T1 H200 + 4-card drafter 513
- page:fourcard-4c1-wave-e-silicon-20261022 | 2026-10-22 | current | Track 1: 4C1 wave E on silicon. ATT 0xE4B712B4 + MLPV 0xE4B712E4 measured ladder, the 4-card archite
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-wave-e-silicon-20261022/
  State: wave E is RUNNING on the 4C1 quad (ATT 0xE4B712B4 md5 a968a8ee on 01/42, MLPV 0xE4B712E4 md5 99f04a91 on 61/62), flashed 2026-10-22 02:5x EDT, cold AC#28 at 03:18. New in this version: the whol
- page:t2-talon-8card-target-20261022 | 2026-10-22 | current | Track 2: 8-card talon target, Qwen done-right on silicon, HBM bandwidth, SLR2 (2026-10-22)
  https://hardware.research.yshicarrick.com/p/t2-talon-8card-target-20261022/
  Done-right Qwen image 0xE4B7126D-CPW (corner-prohibit re-place) was flashed p0 x8 on talon and tested after a cold AC. It runs Qwen3.8-27B TP8 with on-card mixers at 41-42 tok/s batch-1 (2.2x the v8 h
- page:eagle4-tested-data-20261021 | 2026-10-21 | superseded | EAGLE4 all tested data, 2026-10-21
  https://hardware.research.yshicarrick.com/p/eagle4-tested-data-20261021/
  All measured [M] results of Track 1 (H200 + 4-card FPGA drafter, lab-host), Track 2 (8-card talon targets) and Track 3 (constant fold on 8 cards, talon), including the images that failed on silicon. Batch
- page:eagle4-progress-projection-20261021 | 2026-10-21 | superseded | EAGLE4 progress and next 2 days
  https://hardware.research.yshicarrick.com/p/eagle4-progress-projection-20261021/
  What is building or queued on the three tracks, the gate each image must pass, ETAs for 10-21 to 10-23, and the projected batch-1 tok/s if the gate passes. Measured results are on the tested-data post
- page:eagle4-track1-20261021 | 2026-10-21 | current | Track 1: H200 + 4-card FPGA drafter, 2026-10-21
  https://hardware.research.yshicarrick.com/p/eagle4-track1-20261021/
  Track 1 only: EAGLE4 drafter on 4 V80 cards in lab-host verified by a Llama-3B target on the H200. Measured results and next-2-day plan. lab-host data pending: rows come from the delivery notes, not lab-host's ow
- page:eagle4-track2-20261021 | 2026-10-21 | current | Track 2: HBM Qwen 27B + URE drafter + Llama/Qwen on talon, 2026-10-21
  https://hardware.research.yshicarrick.com/p/eagle4-track2-20261021/
  Track 2 only: whole Llama and Qwen models on the 8 V80 cards of talon (TP8, HBM-hosted weights), plus the URE-hosted drafter. Measured results, failures and next-2-day plan. || Qwen3.8-27B 21.1 tok/s 
- page:eagle4-track3-20261021 | 2026-10-21 | current | Track 3: FOLDRING8 constant-folded model on the 8-card ring, 2026-10-21
  https://hardware.research.yshicarrick.com/p/eagle4-track3-20261021/
  Track 3 only: ternary weights folded into LUT logic over the 8 V80 cards of talon. Measured results, the 0x12F8 failure and the next-2-day plan. || 5,845 tok/s batch-1, TriLM-99M folded on 8 V80, bit-
- page:att-ladder-20261020 | 2026-10-20 | current | 4C1 ATT ladder: R3_EARLY, NORMFUSE, G5 on 132 MHz
  https://hardware.research.yshicarrick.com/p/att-ladder-20261020/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), ATT card (0x1288 lineage). The ladder takes cycles out of the ATT item in measured steps: R3_EARLY and NORMFUSE shorten G7G, G5 moves onto the 
- page:fourcard-4c1-chain-cut-20261020 | 2026-10-20 | current | 4C1 chain cut: the level chain as straps (ATT 0xE4B712C6 / MLPV 0xE4B712C7)
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-chain-cut-20261020/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), H200 verifies. The w1 d7 level is 310 us and 107 us of it sits in on-card states and hops without a phase counter (CHAIN_CENSUS_20261020). Each
- page:fourcard-4c1-freeleaf-v2-20261020 | 2026-10-20 | current | 4C1 free leaves v2: depth-1 (host) and every level (RTL)
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-freeleaf-v2-20261020/
  Program 1: the 4-card Llama-3.2-3B drafter (4C1); the H200 verifies. Free leaves fill TREE_STOP's phantom verify slots with the drafter's rank-3+ children; V stays 20. The depth-1 form runs one extra 
- page:fourcard-4c1-att-g5kvc-20261020 | 2026-10-20 | current | KV commit + G5FAST in one ATT image: 0xE4B712BB
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-att-g5kvc-20261020/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter, ATT card. Two ATT features had each passed the QM timing bar: KV commit (0x12B1 / 0x12B2: the accepted path's drafted K/V is copied on card, no per-s
- page:4c1-gpu-ref-humaneval-vs-mtbench-20261020 | 2026-10-20 | current | GPU reference accept grid, humaneval vs mt_bench
  https://hardware.research.yshicarrick.com/p/4c1-gpu-ref-humaneval-vs-mtbench-20261020/
  Program: 4-card Llama-3.2-3B drafter (4C1). GPU H200 exec-mode -1 (bitnet_721, SGLang 0.5.16) over W1-5 x d7/9/10/12/15 on humaneval (80 req) and mt_bench (50 req), 32 new tokens, fa3 and fa4. Grid pa
- page:qwen27b-drafter-serve-accept-20261020 | 2026-10-20 | current | Qwen3.8-27B + BitNet/SubLN EAGLE3 drafter on the H200: serving and accept
  https://hardware.research.yshicarrick.com/p/qwen27b-drafter-serve-accept-20261020/
  Program: the 3-card drafter program re-targeted at Qwen3.8-27B (the TARGET runs on the H200; the drafter is destined for the V80s). This page is the GPU-only step. Current state: the step-21491 drafte
- page:fourcard-4c1-att-g5fast-20261019 | 2026-10-19 | current | 4C1 ATT G5 attention: 16 lockstep head engines halve the per-context cost
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-att-g5fast-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), H200 verifies. On the wave-A ATT silicon census (0x1288), G5 attention is the largest ATT phase: 1,748 of a 6,680-cycle item. The engine runs 3
- page:fourcard-4c1-adaptive-tree-20261020 | 2026-10-20 | current | 4C1 adaptive tree, free leaves and hedge: offline replays
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-adaptive-tree-20261020/
  Program 1: the 4-card Llama-3.2-3B drafter (4C1); the H200 verifies. Offline replays (no card, no GPU) on the card's own TREE_STOP=0 capture (406 w2 d7 trees, mt_bench 50x32) and the GPU drafter's top
- page:qwen27b-4c1-host-converter-20261020 | 2026-10-20 | current | Qwen3.8-27B drafter on the 4C1 pair: converter, card-numerics golden, host profile
  https://hardware.research.yshicarrick.com/p/qwen27b-4c1-host-converter-20261020/
  Program: dual-model 4C1 drafter (Llama-3.2-3B EAGLE4 + Qwen3.8-27B EAGLE3 BitNet/SubLN on one ATT 0xE4B71288 + MLPV 0xE4B7128B pair). The BitNet bf16 export is converted to per-rank card streams by re
- page:fourcard-4c1-verify-kernels-20261020 | 2026-10-20 | current | 4C1 verify kernels K3/K1/K2/K4 at M<=20
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-verify-kernels-20261020/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1); the H200 runs the fp16 target verify. Four verify-graph kernels behind default-off flags (EAGLE4_4C1_VK1..4), each with a NaN-poisoned micro ga
- page:fourcard-4c1-dual-att-20261020 | 2026-10-20 | current | One ATT image for two drafters: Llama-3.2-3B and Qwen3.8-27B by register
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-dual-att-20261020/
  Fork fourcard_4c1_att_dual_20261020 = the 0x12B2 ATT source (0x1288 + KV_COMMIT) plus strap DUAL_MODEL_V. At strap 1 the image comes up as wave B (Llama) and switches to the Qwen drafter by writing MO
- page:qwen27b-4c1-uram-headroom-retrains-20261020 | 2026-10-20 | current | 4C1 headroom for bigger drafter retrains
  https://hardware.research.yshicarrick.com/p/qwen27b-4c1-uram-headroom-retrains-20261020/
  Read-only census of what the two 4C1 images (ATT 0xE4B71288, MLPV 0xE4B7128B r2) hold in URAM, BRAM and HBM, per card and per SLR, from the placed utilization reports and the synthesis RAM mapping. Pu
- page:fourcard-4c1-kvcommit-image-20261020 | 2026-10-20 | current | KV commit in an image: ATT 0xE4B712B2
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-kvcommit-image-20261020/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter. The ATT card already holds the drafted K/V of every forwarded tree node at the RoPE angle of its committed position; COMMIT_PATH (CSR 0x04/0x0C) copi
- page:fourcard-4c1-waveb-deep-20261020 | 2026-10-20 | current | 4C1 wave B deep sweep d8-d15
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-waveb-deep-20261020/
  Program 1: 4-card drafter, wave B rebooted armed at depth 15, measure-only. One round per cell. || w2 d9 V20 525.9 tok/s @ 3.97 (1 round); w2 plateau d8-d10, cards hidden (margin 2.6-2.9 ms); w3 deep 
- page:fourcard-4c1-waveb-sweep-20261020 | 2026-10-20 | current | 4C1 wave B depth x width sweep
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-waveb-sweep-20261020/
  Program 1: 4-card drafter, wave B, depth 2-7 x width 1-4, record-default launcher, measure-only. Cards faulted at w4 d5 stepprof run; cold AC needed. || Best cell w2 d7 V20 524.9 tok/s @ 3.86 (4 round
- page:fourcard-4c1-mlpv-g11glm-v2-20261020 | 2026-10-20 | current | 4C1 MLPV G11+GLM v2: GLM 256 lanes + SUMSQ fold, 1,658 -> 489 cycles
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-mlpv-g11glm-v2-20261020/
  Program: 4-card Llama-3.2-3B EAGLE4 DRAFTER (4C1), MLPV cards. The MLPV item spends G11 RMSNorm + GLM low-rank projection cycles after the G9 exchange. This page measures two follow-on levers in the f
- page:fourcard-4c1-chain-census-20261020 | 2026-10-20 | current | 4C1 chain census: where the w1 level goes
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-chain-census-20261020/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter. Cards ATT 0xE4B71288 x2 + MLPV 0xE4B7128B x2, depth-7 PERSIST, loop modes eager+early_a. Task 1: FSM-mirror (0x080) + loop-counter sampling of 40 tre
- page:fourcard-4c1-waveb-measure-20261020 | 2026-10-20 | current | 4C1 wave B measured: MLPV v1.2
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-waveb-measure-20261020/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter. Wave B = ATT 0xE4B71288 x2 + MLPV v1.2+RSTSYNC 0xE4B7128B x2 (PDI f4f3f745), depth-7 PERSIST. Measure-only: golden bf16-only, inconclusive. Box load 
- page:fourcard-4c1-i1-verify-census-20261019 | 2026-10-19 | current | 4C1 I1: verify CUDA graph census and VQSTRIDE
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-i1-verify-census-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1); the H200 runs the Llama-3.2-3B fp16 target verify. torch.profiler/CUPTI census of 20 verify CUDA graphs (w2 d7, VTOK 20, SGLang 0.5.16, fa4 dra
- page:fourcard-4c1-d1-hosttime-20261019 | 2026-10-19 | current | 4C1 D1 host time: lean attention breaks (T) + cuBLAS pre-warm (D)
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-d1-hosttime-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), H200 verifies. D1 runs the next request's target extend (T) and drafter extend (D) in the current request's step gaps; this page cuts their hos
- page:fourcard-4c1-thin-margin-20261019 | 2026-10-19 | current | 4C1 thin-margin regions: why every build ends at +0.00x and the margin builds launched
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-thin-margin-20261019/
  Program: 4-card Llama-3.2-3B EAGLE4 DRAFTER (4C1). The agent read the intermediate and final timing of every in-flight 4C1 build, opened the routed or post-place checkpoints on a spare host (read-only
- page:fourcard-4c1-dw-host-20261019 | 2026-10-19 | current | 4C1 direct write: KV rows and root input written straight into the ATT card
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-dw-host-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), target verify on the H200. The consolidated ATT image 0xE4B712D1 has an AXI4 slave (s_axi_dw, 1 MiB at 0x0201_0010_0000) that writes KV rows in
- page:fourcard-4c1-record-cell-20261019 | 2026-10-19 | current | 4C1 record cell: every lever on
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-record-cell-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1, wave A ATT 0xE4B71288 + MLPV 0xE4B71287), H200 verifies. w2 d7, VTOK 20, mt_bench 50x32, fa4, theta -2.5, 6 rounds per arm in rotated order. Mea
- page:4c1-rtl-ref-waveab-20261019 | 2026-10-19 | current | 4C1 exact 2-bit references for wave A and wave B (own-RTL co-sim)
  https://hardware.research.yshicarrick.com/p/4c1-rtl-ref-waveab-20261019/
  Program: 4-card Llama-3.2-3B drafter (4C1). The bf16 golden fails at depth>=1 because the 2-bit model differs from bf16, so correctness is judged by bit-identity with a co-simulation of the image's ow
- page:fourcard-4c1-verify-pre-20261019 | 2026-10-19 | current | 4C1 VERIFY_PRE: GPU verify pre-enqueued during the card wait
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-verify-pre-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), H200 verifies. The card process writes the tree's func_info into CUDA-mapped pinned memory and then a seq word; the scheduler has already enque
- page:fourcard-4c1-tree-stop-theta-20261019 | 2026-10-19 | current | 4C1 TREE_STOP theta: accept versus step time on silicon
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-tree-stop-theta-20261019/
  Program: 4-card Llama-3.2-3B drafter (4C1), H200 verifies. The host TREE_STOP mass rule stops the card tree early; the accept walk showed it costs 0.295 accept per step. Here theta is swept on silicon
- page:talon-0xe4b71262-speeds-20261018 | 2026-10-18 | current | talon 8x V80 image 0xE4B71262: served-model speeds
  https://hardware.research.yshicarrick.com/p/talon-0xe4b71262-speeds-20261018/
  Every native model we can serve on talon's 8 V80 cards (6 for TP6 and DeepSeek) on image 0xE4B71262, greedy batch-1, VPU_MODE CSR 0x048=1, beacon off, 3 repeats each. All numbers are measured on silic
- page:mlpv-g11glm-20261019 | 2026-10-19 | current | MLPV G11 + GLM cut: final norm and low-rank lm-down 1,656 -> 692 cycles, bit-exact
  https://hardware.research.yshicarrick.com/p/mlpv-g11glm-20261019/
  Program 1 (4-card Llama-3.2-3B drafter, H200 verifies). On the MLPV card the final RMSNorm (G11) and the low-rank lm-down GEMV (GLM) took 1,656 of ~7,460 listed cycles per item. They ran fully serial 
- page:fourcard-accept-walk-20261019 | 2026-10-19 | current | 4-card drafter accept walk: card trees vs GPU trees, step by step
  https://hardware.research.yshicarrick.com/p/fourcard-accept-walk-20261019/
  Program: 4-card Llama-3.2-3B DRAFTER (4C1), H200 verifies. A new off-by-default capture (EAGLE4_4C1_TREECAP) recorded every draft tree the target verified on silicon, with its verify-20 pick and accep
- page:fourcard-4c1-host-glue-cut-20261019 | 2026-10-19 | current | 4-card drafter: cutting the card-process host glue
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-host-glue-cut-20261019/
  Program 1 (4-card Llama-3.2-3B drafter, H200 verify). Two host-only levers in the card-process glue, each env-gated default off, A/B'd on silicon against the best stack (CUDAPACK + Y_DEFER + D1) at w2
- page:fourcard-backend-kc-ab-20261019 | 2026-10-19 | current | 4C1 card A/B: draft backend, first GPU draft and K_C
  https://hardware.research.yshicarrick.com/p/fourcard-backend-kc-ab-20261019/
  Host-only A/B on the flashed 4-card drafter: draft attention backend (fa3/fa4), first-GPU-draft (lever Y) on/off, and the shortlist size K_C. Six arms, 3 rounds each, mt_bench 50 x 32, w2 d7, 20 verif
- page:fourcard-4c1-d1-stream-20261019 | 2026-10-19 | current | 4C1 D1_STREAM: next-request prefill on a second CUDA stream in the verify window
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-d1-stream-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), H200 verifies. D1 runs the next request's prefill phases (T target extend, D drafter extend) in the card window of the running request. D1_STRE
- page:fourcard-accept-gap-split-20261019 | 2026-10-19 | current | 4-card drafter accept gap 3.65 vs 4.04, split on the GPU
  https://hardware.research.yshicarrick.com/p/fourcard-accept-gap-split-20261019/
  Program: 4-card Llama-3.2-3B DRAFTER (4C1), H200 verifies. The 4-card run accepts 3.65 tokens/step vs 4.04 for the GPU-only drafter (mt_bench 50x32, w2 d7, verify 20). Each card numeric was emulated i
- page:fourcard-kv-prompt-only-20261019 | 2026-10-19 | current | Prompt-only K/V import: card-computed and drafted K/V for accepted tokens
  https://hardware.research.yshicarrick.com/p/fourcard-kv-prompt-only-20261019/
  Program 1 (4-card Llama-3.2-3B drafter, ATT 0xE4B71288 + MLPV 0xE4B71287, H200 target). User idea: the cards should get only the system-prompt K/V from the GPU, not per-step K/V. (A) prompt-only KVI +
- page:fourcard-4c1-d1-prefill-overlap-20261019 | 2026-10-19 | current | 4C1 host levers: next-request prefill in the card gaps, deferred KV import, prefix-resident KV
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-d1-prefill-overlap-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1) with the H200 as target. Three host-only, env-gated levers on the per-request overhead: D1 runs the next queued request's prefill while the card
- page:fourcard-4c1-vs-2card-20261019 | 2026-10-19 | current | 4-card drafter vs 2-card image: end-to-end, raw drafter and per-phase [0xE4B71199,0xE4B711BD]
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-vs-2card-20261019/
  Silicon measurements of the 4-card EAGLE4 drafter (2 attention cards + 2 MLP/lm-head cards, tensor-parallel by 2, H200 target verify) against the 2-card record image: an end-to-end tree-shape sweep, t
- page:fourcard-4c1-ctrlfsm-20261019 | 2026-10-19 | current | 4C1 control FSM: one doorbell per card replaces the per-step host register choreography
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-ctrlfsm-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1), target verify on the H200. Measured wave-A step at w2 d7: p50 7.08 ms, of which GPU verify wait 3.81, card tree loop 1.29 and host glue about 2
- page:g5-ctx-scaling-4c1-att-20261019 | 2026-10-19 | current | G5 attention context scaling on the 4C1 ATT card
  https://hardware.research.yshicarrick.com/p/g5-ctx-scaling-4c1-att-20261019/
  Program: 4C1 3-card/4-card Llama-3.2-3B DRAFTER, ATT pair. Measured the G5 cycle law of the 0xE4B71288 ATT engine in xsim and prototyped HEAD_PAR 16 / G5_HPS 1 in fork fourcard_4c1_g5scale_20261019. N
- page:rca-att-g4-silicon-vs-xsim-20261019 | 2026-10-19 | current | ATT G4: silicon 1,074-1,097 vs xsim 1,843 cycles
  https://hardware.research.yshicarrick.com/p/rca-att-g4-silicon-vs-xsim-20261019/
  The G4 counter on 4C1 v0 ATT reads 1,074-1,097 cycles per item on silicon against 1,843 in xsim. The cards run G4_EMBFOLD=1; the xsim ran it off. With the fold on, the same RTL gives 1,027 per item. |
- page:fourcard-onchip-prefill-wcol-20261019 | 2026-10-19 | current | On-chip prompt K/V with a W-column engine (4C1 ATT)
  https://hardware.research.yshicarrick.com/p/fourcard-onchip-prefill-wcol-20261019/
  Measured the K/V-only G4 pass (12 of 18 tiles) at W=1/4/8 in xsim; per-token input bytes equal today's KVI; fc stays on GPU. || K/V-only sweep costs +3.2% at W=8 (xsim); SLICE binds the fit
- page:fourcard-sibling-per-pair-20261019 | 2026-10-19 | current | One sibling per pair: data-parallel siblings on the 4 cards [0xE4B71199,0xE4B711BD]
  https://hardware.research.yshicarrick.com/p/fourcard-sibling-per-pair-20261019/
  Program 1 (4-card Llama-3.2-3B drafter). Idea: two full record pairs, one W=2 sibling each, instead of tensor-parallel siblings run in series. Read-only: parsed captured GPU trees, the measured step s
- page:fourcard-4c1-host-glue-fsm-20261019 | 2026-10-19 | current | 4C1 host glue into an on-card control FSM
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-host-glue-fsm-20261019/
  Program 1: the 4-card Llama-3.2-3B drafter (4C1). Question: how much of the ~2.1 ms host glue per step can move into the cards, and what P2P paths exist on this box. Source: one wave A w2 d7 profile r
- page:fourcard-4c1-tree-shape-20261019 | 2026-10-19 | current | 4C1 tree shape: which width, depth and verify-token count is fastest
  https://hardware.research.yshicarrick.com/p/fourcard-4c1-tree-shape-20261019/
  Program 1: the 4-card Llama-3.2-3B EAGLE4 drafter (4C1) with the H200 as target. Question: which tree shape maximises tok/s now that a wave-A step is about 9.6 ms with 2.1 to 2.3 ms of host spin on th
- page:talon-ring8-llama-tp8-speeds-20261018 | 2026-10-18 | current | Llama 1B / 3B / 8B on 8 V80 cards (talon, TP8 ring)
  https://hardware.research.yshicarrick.com/p/talon-ring8-llama-tp8-speeds-20261018/
  Llama-3.2-1B, Llama-3.2-3B and Llama-3.1-8B ran tensor-parallel (TP8) on talon's 8 Alveo V80 cards with image 0xE4B71262 (v8), every token bit-exact to the integer simulator. All numbers are batch-1, 
- page:build-reports | 2026-10-18 | current | Build reports (read-only)
  https://hardware.research.yshicarrick.com/p/build-reports/
  Index of read-only build-report browsers; every file linkable by line. || 11 report pages
- page:rpt-1262-v8-s2 | 2026-10-18 | current | Build reports: 1262_v8_s2 (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-1262-v8-s2/
  Read-only copies of the 1262_v8_s2 build reports, linkable by line (?f=<run>/<file>#L219-L230). || 5 files
- page:rpt-1261-v6-s1b | 2026-10-18 | current | Build reports: 1261_v6_s1b (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-1261-v6-s1b/
  Read-only copies of the 1261_v6_s1b build reports, linkable by line (?f=<run>/<file>#L219-L230). || 5 files
- page:rpt-ure-v2e | 2026-10-17 | current | Build reports: ure_v2e (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-ure-v2e/
  Read-only copies of the ure_v2e build reports, linkable by line (?f=<run>/<file>#L219-L230). || 16 files
- page:rpt-1263-v9b-s1 | 2026-10-17 | current | Build reports: 1263_v9b_s1 (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-1263-v9b-s1/
  Read-only copies of the 1263_v9b_s1 build reports, linkable by line (?f=<run>/<file>#L219-L230). || 3 files
- page:ring8-timing-diagnosis-20261017 | 2026-10-17 | current | Ring-8 Timing Diagnosis
  https://hardware.research.yshicarrick.com/p/ring8-timing-diagnosis-20261017/
  v8 0xE4B71262 (C_COLS 2 @225, GT island, crumbc fanout) closed timing (+0.003 / true +0.093, 0 failing) and passes on silicon: HBMS ladder 52/52 on all 8 talon cards, ring BIST 16/16 PASS 0 errors || 
- page:rpt-1262-v8-s1 | 2026-10-17 | current | Build reports: 1262_v8_s1 (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-1262-v8-s1/
  Read-only copies of the 1262_v8_s1 build reports, linkable by line (?f=<run>/<file>#L219-L230). || 4 files
- page:rpt-125f-kitb | 2026-10-17 | current | Build reports: 125F_kitB (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-125f-kitb/
  Read-only copies of the 125F_kitB build reports, linkable by line (?f=<run>/<file>#L219-L230). || 122 files
- page:rpt-125f-kitbm | 2026-10-17 | current | Build reports: 125F_kitBM (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-125f-kitbm/
  Read-only copies of the 125F_kitBM build reports, linkable by line (?f=<run>/<file>#L219-L230). || 121 files
- page:rpt-council-2026-10-17 | 2026-10-17 | current | Council results (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-council-2026-10-17/
  Read-only copies of the council-2026-10-17 build reports, linkable by line (?f=<run>/<file>#L219-L230). || 8 files
- page:track1-4c1-engine-floorplan-20261017 | 2026-10-17 | current | Track 1: 4C1 engine floorplan per card (failed timing builds)
  https://hardware.research.yshicarrick.com/p/track1-4c1-engine-floorplan-20261017/
  ATT and MLPV failed only on the link clock; MLPV u_tree_level +0.003 ns misses the +0.03 release gate; ATT GEMV array +0.234 ns is its tightest drafter path. || Clickable engine floorplans of both fai
- page:rpt-125f-s1c | 2026-10-17 | current | Build reports: 125F_s1c (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-125f-s1c/
  Read-only copies of the 125F_s1c build reports, linkable by line (?f=<run>/<file>#L219-L230). || 129 files
- page:rpt-1256-s1r | 2026-10-17 | current | Build reports: 1256_s1r (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-1256-s1r/
  Read-only copies of the 1256_s1r build reports, linkable by line (?f=<run>/<file>#L219-L230). || 127 files
- page:rpt-1254-s2-good | 2026-10-17 | current | Build reports: 1254_s2_GOOD (2026-10-17)
  https://hardware.research.yshicarrick.com/p/rpt-1254-s2-good/
  Read-only copies of the 1254_s2_GOOD build reports, linkable by line (?f=<run>/<file>#L219-L230). || 72 files
- page:track1-4c1-att-failed-build-per-phase-20261017 | 2026-10-17 | current | Track 1: 4C1-ATT per-phase architecture, read from the failed timing build only
  https://hardware.research.yshicarrick.com/p/track1-4c1-att-failed-build-per-phase-20261017/
  Program: 3B Llama-3.2 2-bit EAGLE4 drafter, 4-card 4C1, attention card (cards 01 + 42). This page uses ONE build only: the failed-timing ATT build s3LAattf2 on lab-host (route WNS -0.769 ns). It shows eac
- page:track1-4card-drafter-architecture-shift-20261016 | 2026-10-16 | current | Track 1: the 4-card drafter architecture shift (record pair to 4C1) [0xE4B71199,0xE4B711BD]
  https://hardware.research.yshicarrick.com/p/track1-4card-drafter-architecture-shift-20261016/
  Track 1 = the 3B Llama-3.2 2-bit EAGLE4 DRAFTER on V80 cards (target on H200), batch 1, goal: beat 350.16 tok/s. This page shows, read from the RTL, what changes between the record 2-card pair (HEAD 0
- page:fourcard-drafter-slicing-20261016 | 2026-10-16 | current | 4-card drafter: where the time goes and how to slice the weights across 4 V80s [0xE4B71199,0xE4B711BD]
  https://hardware.research.yshicarrick.com/p/fourcard-drafter-slicing-20261016/
  Program: the 3B Llama-3.2 2-bit EAGLE4 DRAFTER on V80 cards (target on H200), batch 1. This page pins what the record 2-card pair (HEAD 0xE4B711BD + TAIL 0xE4B71199, 350.16 tok/s @ 3.65) actually spen
- page:ternary-make-possible-20261011 | 2026-10-11 | current | Ternary constant folding on 8 V80s: measured levers, full reproduction pack (Track T)
  https://hardware.research.yshicarrick.com/p/ternary-make-possible-20261011/
  <b>Program 2, Track T. Owner: lab-host-golden-laptop.</b> <b>State:</b> a 2-layer TriLM-99M card (single image, weights loaded at runtime) closes 400 MHz (WNS +0.057, TNS 0, WHS +0.011) in ONE SLR after r
- page:eagle4-v80-platform | 2026-10-15 | current | AMD Alveo V80: our platform
  https://hardware.research.yshicarrick.com/p/eagle4-v80-platform/
  One Versal HBM device (xcv80, VH1782) with three SLRs, 32 GB HBM2e, four QSFP cages and a PCIe Gen5 x8 host link; we run 12 of them across two machines. || Clickable V80: 27 topics (SLRs, HBM, PCIe, Q
- page:eagle4-model-weights | 2026-10-15 | current | EAGLE4 Model Weights
  https://hardware.research.yshicarrick.com/p/eagle4-model-weights/
  Pick a model and a tensor to see its exact shape, bit width, size and placement, with the related research posts below. || 7 models, 76 tensors: exact shapes, bit widths and where each lives
- page:talon-serving-llama-20261015 | 2026-10-15 | current | Serving Llama on talon V80s with HBM streaming: 1B to 70B, no GPU
  https://hardware.research.yshicarrick.com/p/talon-serving-llama-20261015/
  Real Hugging Face Llama models were compiled into the HBMS engine's descriptor program (embedding, all layers, LM head and argmax on card) and run batch-1 on talon's V80s with the already-flashed imag
- page:eagle4-arch-book | 2026-10-15 | current | EAGLE4 Architecture Book [0xE4B71197,0xE4B71199,0xE4B7119F,0xE4B711A3...]
  https://hardware.research.yshicarrick.com/p/eagle4-arch-book/
  Pick a track and an image to see its block diagram, what it added, its timing and its measured tok/s. || 43 images across 3 tracks, one interactive block diagram per image
- page:eagle4-knowledge-base | 2026-10-15 | current | EAGLE4 Knowledge Base [0xE4B71111,0xE4B71112,0xE4B71113,0xE4B71119...]
  https://hardware.research.yshicarrick.com/p/eagle4-knowledge-base/
  The pinned knowledge graph plus a searchable index of every HQ report, repo report and build script, the findings ledger, branches and the image registry. || 17,427 files · 1,269 findings · 190 branch
- page:talon-8card-serving-20261015 | 2026-10-15 | current | Talon 8-card serving: one model across 8 V80s
  https://hardware.research.yshicarrick.com/p/talon-8card-serving-20261015/
  Track 2 sprint: serve one model across all 8 talon V80 cards. Single-card serving and the ring links are proven on silicon; the 8-card split is compiled and bit-exact in emulation, waiting for the ima
- page:talon-hbms-engine-arch-cycles-20261015 | 2026-10-15 | current | Talon V80 HBMS engine: architecture, Vivado results and silicon cycle profile
  https://hardware.research.yshicarrick.com/p/talon-hbms-engine-arch-cycles-20261015/
  Bookkeeping page for the talon 8x V80 HBMS-L engine (image 0xE4B71254 s2): what the engine is, what Vivado placed and routed, and where every cycle of a decoded token goes on silicon. Llama-3.2-1B and
- page:eagle4-build-board | 2026-10-15 | current | EAGLE4 unified build board
  https://hardware.research.yshicarrick.com/p/eagle4-build-board/
  One control point (user, 2026-10-11: "now talk to hq for unified control"). HQ holds the board: hq-junior-2 for RING8 / 8-card architecture and the board, HQ main for operations (mesh, QM, cold-AC con
- page:eagle4-tunit-constant-fold | 2026-10-14 | current | Constant-fold ternary unit on one V80: TriLM-99M layer 0, bit-exact on silicon
  https://hardware.research.yshicarrick.com/p/eagle4-tunit-constant-fold/
  A ternary (−1/0/+1) LLM layer folded into V80 LUTRAM: weights are written at runtime, so the same bitstream serves any layer of this shape on any card. Layer 0 of TriLM-99M (q, k, v, o, gate, up, down
- page:eagle4-memory-biosystem | 2026-10-14 | current | A living memory for agent sessions: dawn/day/dusk dynasties and a Haiku dreamer
  https://hardware.research.yshicarrick.com/p/eagle4-memory-biosystem/
  Our FPGA program runs many long-lived Claude Code sessions. Their history grows without bound (5.69 GB), new sessions inherited rules without reasons, and the user had to re-teach operations again and
- page:eagle4-collaboration-org | 2026-10-14 | current | EAGLE4 collaboration: who runs what, who reports to whom
  https://hardware.research.yshicarrick.com/p/eagle4-collaboration-org/
  Who is in charge of what across build-host-A (HQ) and lab-host (lab), and the working rules between them. Roster as of 2026-10-14 06:41 EDT. The user decides only cold AC, sudo, deletions, permission rules, sto
- page:spec-decode-eagle4-vs-dmatrix | 2026-10-14 | current | Speculative decoding: our EAGLE4 FPGA drafter vs d-Matrix Corsair [0xE4B71143]
  https://hardware.research.yshicarrick.com/p/spec-decode-eagle4-vs-dmatrix/
  Comparison for the 3-card/2-card Llama-3.2-3B DRAFTER program (program 1; the 8-card RING8 full-target program is noted separately). Research motto: a track never ends by denying the idea; walls are m
- page:lab-host-qwen-quant-gate-20261011 | 2026-10-11 | current | Qwen3.6-27B quantization + accuracy gate (W8 / W4 / W3 / W2, contract formats)
  https://hardware.research.yshicarrick.com/p/lab-host-qwen-quant-gate-20261011/
  <b>Question (USER, via HQ):</b> "The line it serves is Qwen3.6-27B, does it have 2-bit weights though?" <b>Answer:</b> no — Qwen3.6-27B ships bf16 only (snapshot 6a9e13bd, 52 GB); no W2/W3/ternary che
- page:lab-host-drafter-int8act-20261011 | 2026-10-11 | current | 8-bit activations in the Llama-3.2-3B EAGLE4 drafter (Program 1)
  https://hardware.research.yshicarrick.com/p/lab-host-drafter-int8act-20261011/
  CURRENT STATE (v3, 2026-10-11 03:55): Program 1 (3-card Llama-3.2-3B EAGLE4 drafter). The six SGLang arms ran on the H200 on the pinned GPU accept command (humaneval 80 requests, depth 15, width 4, 64
- page:lab-host-golden-suite-20261011 | 2026-10-11 | current | Golden captures and RTL verification suites (BitNet-2B-4T, Qwen3.6-27B, EAGLE4 drafter)
  https://hardware.research.yshicarrick.com/p/lab-host-golden-suite-20261011/
  CURRENT STATE (v5, 2026-10-11 04:05 data): the shared verification suite for both programs (Program 1 = 3-card Llama-3.2-3B EAGLE4 drafter; Program 2 = 8-card full target). One HF capture tool, 5 Tier
- page:eagle4-knowledge-graph | 2026-10-07 | current | EAGLE4 knowledge graph
  https://hardware.research.yshicarrick.com/p/eagle4-knowledge-graph/
  Discovery layer over every page, report, note and skill, linked to images and features; Images carry post-route timing, utilization, verdicts and findings from the HQ export; lineage from HQ parent ed
- page:agentic-session-history | 2026-10-11 | current | Agentic session history [0xE4B71199,0xE4B711BD]
  https://hardware.research.yshicarrick.com/p/agentic-session-history/
  Curated, scrubbed history of the EAGLE4 agent sessions (319 sessions/subagents, 93 events) from HQ's on-disk records; calculations in the top bar, recorded facts below; HQ re-drops it as the roster ch
- page:lab-host-osprey-goldens-20261011 | 2026-10-11 | superseded | OSPREY per-op goldens + 3STAGE seam vectors (Program 1)
  https://hardware.research.yshicarrick.com/p/lab-host-osprey-goldens-20261011/
  SUPERSEDED: this content now lives in the golden-suite page lab-host-golden-suite-20261011 (one-page rule). The 3STAGE seam contract was later PINNED by HQ (h1 Q11.5, G9lo Q8.8, natural lane order, lane-c
- page:ring8-hbm-llm-serving-20261011 | 2026-10-11 | current | 8-FPGA HBM ring LLM serving with a model contract
  https://hardware.research.yshicarrick.com/p/ring8-hbm-llm-serving-20261011/
  CURRENT STATE (v3, 2026-10-11 ~03:30 EDT): umbrella page of the 8-card FULL TARGET program (not the 3-card EAGLE4 drafter). Two image tracks: Track T = ternary constant folding (page ternary-make-poss
- page:qwen27b-ring8-20261011 | 2026-10-11 | current | Track Q: Qwen3.6-27B on 8 V80s (HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA)
  https://hardware.research.yshicarrick.com/p/qwen27b-ring8-20261011/
  CURRENT STATE (v3, 2026-10-11 ~03:30 EDT): REOPENED as the one living page of Track Q of the 8-card FULL TARGET program (user order: "HBM-maximized-flux Qwen3.6-27B on 8 cards with QSFP ring + PCIe FP
- page:lab-host-qwen-hbm-flux-8card-20261011 | 2026-10-11 | superseded | Track Q: Qwen3.6-27B on 8 FPGAs — HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA
  https://hardware.research.yshicarrick.com/p/lab-host-qwen-hbm-flux-8card-20261011/
  SUPERSEDED 2026-10-11: merged into the one living page for this topic, qwen27b-ring8-20261011 (https://hardware.research.yshicarrick.com/p/qwen27b-ring8-20261011/). Body below kept as the record of th
- page:lab-host-axi-testbench-20261011 | 2026-10-11 | superseded | Golden AXI suite: protocol checkers, BFMs, AXI input testbench
  https://hardware.research.yshicarrick.com/p/lab-host-axi-testbench-20261011/
  SUPERSEDED 2026-10-11: merged into the one living page for this topic, lab-host-golden-suite-20261011 (https://hardware.research.yshicarrick.com/p/lab-host-golden-suite-20261011/). Body below kept as the reco
- page:ternary-constfold-20261011 | 2026-10-11 | superseded | Ternary constant-folded LUT adder trees (BitNet/TriLM) [0xE4B71128]
  https://hardware.research.yshicarrick.com/p/ternary-constfold-20261011/
  SUPERSEDED 2026-10-11: merged into the one living page for this topic, ternary-make-possible-20261011 (https://hardware.research.yshicarrick.com/p/ternary-make-possible-20261011/). Body below kept as 
- page:lab-host-ternary-fold-8card-20261011 | 2026-10-11 | superseded | Track T: constant folding on 8 FPGAs — TriLM-99M fully folded + a co-resident folded drafter
  https://hardware.research.yshicarrick.com/p/lab-host-ternary-fold-8card-20261011/
  SUPERSEDED 2026-10-11: merged into the one living page for this topic, ternary-make-possible-20261011 (https://hardware.research.yshicarrick.com/p/ternary-make-possible-20261011/). Body below kept as 
- page:eagle4-4card-arch-variants-20261007 | 2026-10-07 | current | EAGLE4 on Four V80s: Variants and Choices
  https://hardware.research.yshicarrick.com/p/eagle4-4card-arch-variants-20261007/
  Diff from the 10-05 architecture post: what each 4-card variant became, the measured verdicts, and why CL4 ring + shape-generic + HBM streaming are the next bets. || SIBX-E 4-card rejected (X 114.35 v
- page:eagle4-trace-20261006 | 2026-10-06 | current | EAGLE4 Trace: per-block timelines, bubbles and utilisation
  https://hardware.research.yshicarrick.com/p/eagle4-trace-20261006/
  Perfetto traces per hardware block from xsim + lab-host hostprof; the draft step is a serial host->cards->GPU chain, so overlap beats engine speed-up. || Silicon step: HEAD card busy 5-17 %, top bubble = 
- page:eagle4-arch-evolution-20261005 | 2026-10-05 | current | EAGLE4 Architecture Evolution
  https://hardware.research.yshicarrick.com/p/eagle4-arch-evolution-20261005/
  Animated walk-through of the EAGLE4 V80 drafter: HEAD/TAIL split, 256-lane DSP58 GEMV, KV import, LEVEL_PIPE/FG8, 75 MHz TAIL record 339.12 tok/s [M], QSFP D2/CL4 4-card paths [P], and runtime shapes.
- page:eagle4-roadmap-review-pack-20261005 | 2026-10-05 | current | EAGLE4 Roadmap Review Pack
  https://hardware.research.yshicarrick.com/p/eagle4-roadmap-review-pack-20261005/
  Curated pages and files to review with partner hardware leads, grouped by the roadmap decision each one feeds. || Reading list for the roadmap session: 6 themes, site pages + lab files, with discussio
- page:eagle4-weekly-arch-20261005 | 2026-10-05 | current | EAGLE4 V80 Architecture Week 40 [0xE4B71197,0xE4B71199,0xE4B711BD,0xE4B711EA]
  https://hardware.research.yshicarrick.com/p/eagle4-weekly-arch-20261005/
  Drafter split across two V80s: HEAD = embedding, QKV, attention, O, MLP gate/up on a 128-lane GEMV array; TAIL = SiLU, down-proj, fc_out, low-rank LM head, vocab tree and G8 rescore. Host relays the s
- page:eagle-companion-chip | 2026-09-30 | current | Eagle Companion Chip
  https://hardware.research.yshicarrick.com/p/eagle-companion-chip/
  On-chip speculative-decoding drafter: 1.15x (MT-Bench) / 1.33x (HumanEval) vanilla H200 at batch 1; FPGA evaluation system, SGLang, switchable target model || 1.33x vanilla H200 on HumanEval, 331 tok/
- page:eagle4-weight-residency | 2026-09-29 | current | EAGLE4 Weight Residency
  https://hardware.research.yshicarrick.com/p/eagle4-weight-residency/
  41.58 MiB 2-bit body on-chip (URAM), fp16 lm-head and embeddings in HBM; 4 V80s hold ~0.9B 2-bit weights, ~5x today || 41.58 MiB on-chip; 4 cards ~5x
- page:eagle4-rtl-onboarding | 2026-09-29 | current | EAGLE4 RTL Onboarding [0xE4B711B4,0xE4B711BD,0xE4B711CC,0xE4B711DB...]
  https://hardware.research.yshicarrick.com/p/eagle4-rtl-onboarding/
  Handbook for new engineers: architecture, repo, toolchain, first build, xsim gates, adjudication and relay, budgets, glossary || Start working on the EAGLE4 RTL
- page:eagle4-week-review-20260927 | 2026-09-27 | superseded | One week on a two-FPGA speculative drafter: what was asked, what got built, and what the silicon sai [0xE4B711B1,0xE4B711BD,0xE4B711BE,0xE4B711BF...]
  https://hardware.research.yshicarrick.com/p/eagle4-week-review-20260927/
  Latest: batch-1 record now 323.30 tok/s (09-29, see the progress page linked below). A review of 2026-09-24 → 09-27 for readers who want to see how the work was actually done. The drafter is a two-car
- page:eagle4-progress-20260929 | 2026-09-29 | superseded | EAGLE4 progress 09-27 → 09-29: new batch-1 record 323.30 tok/s [0xE4B711AF,0xE4B711B4,0xE4B711B7,0xE4B711BB...]
  https://hardware.research.yshicarrick.com/p/eagle4-progress-20260929/
  Two days of silicon and builds after the week review. The two-V80 drafter pair (HEAD + TAIL cards, SGLang EAGLE on one H200) now runs 323.30 tok/s batch-1 on mt_bench w2 d3 (measured, 3 clean rounds, 
- page:eagle4-drafter-journey-20260927 | 2026-09-27 | superseded | Four days on the EAGLE4 FPGA drafter pair: what was slow, why, and what moved
  https://hardware.research.yshicarrick.com/p/eagle4-drafter-journey-20260927/
  A learning write-up of 2026-09-24 → 09-27 on the two-card V80 drafter that feeds SGLang EAGLE on an H200. It walks the system end to end (engines, lanes, bit widths, clocks, bytes per item, URAM and H
- page:tested-images-20260926 | 2026-09-26 | superseded | Tested Images 2026-09-26 [0xE4B711A3,0xE4B711A5,0xE4B711A8,0xE4B711C5...]
  https://hardware.research.yshicarrick.com/p/tested-images-20260926/
  Best 1-round pair: A5+F4 KVI+FASTGLUE4 mt w2 d2 268.01 tok/s @2.55 [measured]; 6 pairs catalogued with w x d sweeps. || Every pair image tested on silicon on 2026-09-26
- page:sglang-draft-kv-lifecycle-20260925 | 2026-09-25 | historical | SGLang draft-KV lifecycle vs the pair [0xE4B711CF,0xE4B711E1]
  https://hardware.research.yshicarrick.com/p/sglang-draft-kv-lifecycle-20260925/
  Read SGLang 0.5.16 CDT's draft-KV lifecycle from source and generalised the calibrated accept emulator to depth-N / width-W trees. The unprimed pair's saturation at depth 5 is reproduced within 0.04 a
- page:host-glue-structural-20260925 | 2026-09-25 | historical | Host glue G: ranked terms and two thread-free / pinned-CPU levers [0xE4B711C8,0xE4B711CF]
  https://hardware.research.yshicarrick.com/p/host-glue-structural-20260925/
  Host glue G is 8.58 ms/step on the pair's default path, which is larger than the whole draft and blocks the bs1 win bar (D_pair + G < 3.42 ms). This page ranks what is left of G, builds the two larges
- page:head-g7seam-pipe-20260924 | 2026-09-24 | historical | G7 seam write pipeline for a faster HEAD clock [0xE4B711E5]
  https://hardware.research.yshicarrick.com/p/head-g7seam-pipe-20260924/
  The 100 MHz fmax census of the routed 0x11C9 HEAD fails 542,754 endpoints; 323,420 of them are the write ports of the two G7 register-file seams (u_seam_r4g / u_seam_r4u), all driven by one write-inde
- page:pair-frspec-prep-20260924 | 2026-09-24 | historical | FR-Spec vocab shortlist on the HEAD/TAIL pair: silicon result [0xE4B71167,0xE4B711C8,0xE4B711CF]
  https://hardware.research.yshicarrick.com/p/pair-frspec-prep-20260924/
  FR-Spec reorders the vocabulary by token frequency so the TAIL scans only the first N of 1,002 tiles. The running TAIL already has the register; this work built the permuted OSPREY weights with gates 
- page:head-prime-batch-20260924 | 2026-09-24 | historical | HEAD PRIME_BURST: many KV-only prime rows per doorbell [0xE4B711E1]
  https://hardware.research.yshicarrick.com/p/head-prime-batch-20260924/
  <p>Priming the HEAD KV with the prompt prefix is the gap between the pair's primed and unprimed runs (primed w3 d2 77.83 tok/s @ accept 2.61, unprimed 146.78 @ 1.97; both <span class="k measured">meas
- page:engine-split-images-20260924 | 2026-09-24 | historical | Engine-split design D: first GEMV-card and state-card images (fit/timing probes) [0xE4B711C8,0xE4B711D0,0xE4B711D1]
  https://hardware.research.yshicarrick.com/p/engine-split-images-20260924/
  Design D splits the drafter by engine across two V80s: a GEMV card with every body GEMV and a state card with attention/KV/G12/TREE/G8, joined by four hops over QSFP. The RTL for both roles and the fr
- page:research-overview-20260924 | 2026-09-24 | historical | EAGLE4 research overview — 2026-09-24 [0xE4B711C8,0xE4B711CD,0xE4B711CF]
  https://hardware.research.yshicarrick.com/p/research-overview-20260924/
  Index of the 19 research-site pages authored or refreshed on 2026-09-24: 14 existing tracks updated to their current state, plus 5 new pages from today's silicon baseline, accept-gap RCA, build-placem
- page:eagle4-restore-fmax-census-20260924 | 2026-09-24 | historical | What stops the whole drafter clock at 100 and 133 MHz (routed HEAD 0xE4B711C9, TAIL 0xE4B711CA) — HE [0xE4B711B0,0xE4B711C9,0xE4B711CA,0xE4B711CD]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-fmax-census-20260924/
  <p>The drafter runs on ONE clock, <code>clock_reset_clk_usr_0</code> = 66.002 MHz (15.151 ns), from the shell clk_wiz (master <code>clk_pl_1</code>, 33.333 MHz). A separate track (array_dualclk) has a
- page:eagle4-restore-array-dualclk-20260924 | 2026-09-24 | historical | 2-bit GEMV array on its own 2x clock (132 MHz array, 66 MHz drafter) — CD ROUTED [0xE4B711C9,0xE4B711CD]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-array-dualclk-20260924/
  The HEAD card sets the pair's pace, and 71.4% of its busy time is the 2-bit GEMV array (G4/G6/G7g/G7u). The whole drafter is capped at 66 MHz by an attention path, but the array alone routes much fast
- page:eagle4-restore-sibling-batching-20260924 | 2026-09-24 | historical | Sibling batching W on the design-D GEMV card: one weight sweep for W children [0xE4B711D3]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-sibling-batching-20260924/
  Every tree level after the root spawns up to 8 independent children that read the SAME 2-bit weights. Sibling batching gives each lane W activation columns so one URAM weight read serves W children. I
- page:eagle4-restore-state-card-20260924 | 2026-09-24 | historical | Engine split, design D steps 1–2: GEMV card (role 3) + state card (role 4) + hop_link, bit-exact in  [0xE4B71155]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-state-card-20260924/
  Design D splits the EAGLE4 drafter across two V80 cards by engine. The GEMV card (PIPE_ROLE 3) runs every 2-bit body GEMV. The state card (PIPE_ROLE 4) owns what is stateful: the attention KV cache, t
- page:eagle4-restore-systolic-lutgemm-20260924 | 2026-09-24 | historical | Systolic activation flow and multiply-free LUT-GEMM on the 2-bit GEMV array (module-only) [0xE4B71155,0xE4B711C9]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-systolic-lutgemm-20260924/
  Two architecture ideas were prototyped on the 2-bit GEMV array alone (the engine behind G4/G6/G7g/G7u/G9/G12), behind parameters whose defaults reproduce today's array cycle-for-cycle. A: activations 
- page:consolidated-images-20260924 | 2026-09-24 | historical | Consolidated next images: HEAD E3 + TAIL 0xE4B711F1 [0xE4B711F1]
  https://hardware.research.yshicarrick.com/p/consolidated-images-20260924/
  The next HEAD/TAIL pair built strictly from changes whose gates are green on the exact RTL. HEAD = E3, already building; TAIL = CF + T1 (3 vocab columns in flight), new build 0xE4B711F1 on build-vm. S1
- page:eagle4-restore-host-overhead-20260924 | 2026-09-24 | historical | Pair host round-trip overhead: waterfall, HOSTFAST, and exact draft/verify overlap [0xE4B711B0,0xE4B711C9]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-host-overhead-20260924/
  The two-card pair (HEAD 0xE4B711C9 + TAIL 0xE4B711B0, depth 2) served 102.13 tok/s while its busiest card sat idle about 40% of each request. A timing-only profiler split every SGLang verify step into
- page:per-op-xval-tooling-20260924 | 2026-09-24 | historical | Per-op GPU-vs-pair drafter cross-validation tool
  https://hardware.research.yshicarrick.com/p/per-op-xval-tooling-20260924/
  A per-op comparator between the GPU (SGLang) reference drafter and the FPGA HEAD/TAIL pair's own captured intermediate values, built to answer whether the pair's much smaller depth-3 accept gain (vs t
- page:mesh-new-build-vms-20260924 | 2026-09-24 | historical | Two new mesh build hosts: Oleg's VM2 and VM3
  https://hardware.research.yshicarrick.com/p/mesh-new-build-vms-20260924/
  Two new virtual machines (named build-vm and build-vm for the mesh) were provisioned end to end as Vivado build hosts: OS packages, 64 GiB swap, a full copy of the Vivado 2025.1 toolchain relayed host-t
- page:build-placement-insights-20260924 | 2026-09-24 | historical | Build placement insights: L256 GEMV card routing, HP8 HEAD timing edge, 100 MHz G7 seam blockers
  https://hardware.research.yshicarrick.com/p/build-placement-insights-20260924/
  A read-only consolidation of every build finished 2026-09-22 through 09-24, extracted from each build's own place/route report with scripts/util_row.py. It answers six placement questions across the H
- page:eagle4-restore-head-levers-20260924 | 2026-09-24 | historical | HEAD levers: 256-bit seam export and KV-only attention skip [0xE4B71137,0xE4B71185,0xE4B711CD]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-head-levers-20260924/
  Two cheap HEAD-card levers from the one-pair win model, implemented as default-off config toggles in a new fork of the 0xE4B711CD HEAD source (verilog 8e6f826). SEAM_WIDE widens the per-child seam exp
- page:eagle4-restore-engine-split-pair-20260924 | 2026-09-24 | historical | Engine split design D, step 3: GEMV card + state card over the real hop link, in one xsim [0xE4B711B0,0xE4B711C9]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-engine-split-pair-20260924/
  <p>Design D splits the drafter into a <b>GEMV card</b> (PIPE_ROLE 3: every 2-bit body GEMV) and a <b>state card</b> (PIPE_ROLE 4: attention, KV, G12/TREE/G8, tree FSM). They exchange 4 hops. Steps 1 a
- page:eagle4-restore-gpu-accept-grid-20260924 | 2026-09-24 | historical | GPU-only accept and tok/s grid: tree width × depth on the pair's checkpoint (H200) [0xE4B711CF]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-gpu-accept-grid-20260924/
  The pair drafts on two V80s and verifies on the H200. The earlier depth-N pair model concluded that tree width 3–4 beats width 8 by +27–35% at depth 2, but it borrowed the accept-vs-width numbers from
- page:pair-accept-gap-priming-20260924 | 2026-09-24 | historical | The pair-vs-GPU accept gap is a priming/regime effect, not a drafter defect [0xE4B711B0,0xE4B711C8,0xE4B711C9,0xE4B711CF]
  https://hardware.research.yshicarrick.com/p/pair-accept-gap-priming-20260924/
  Three read-only tracks tie together tonight: (1) a per-level golden comparison of the pair's on-card tree-decode against a GPU truth tree, which PASSED and rules out an RTL defect in candidate ranking
- page:eagle4-restore-tree-width-20260924 | 2026-09-24 | historical | Runtime tree width on the HEAD/TAIL pair: TAIL image 0xE4B711CF for width 3 — ON THE CARD, silicon s [0xE4B711C8,0xE4B711CB,0xE4B711CF]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-tree-width-20260924/
  The user asked to run the pair at tree width 3 (2026-09-24 01:10). The depth-N model projected +27-35% tok/s over width 8 at depth 2. Width was hard-wired (K_F=8). This track ported the sibling fork's
- page:eagle4-restore-laptop-build-d6-20260924 | 2026-09-24 | historical | Two more PDI build hosts (laptop, build-host-A) and the D6 GEMV-card fit probe [0xE4B711D2,0xE4B711D3,0xE4B711D6]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-laptop-build-d6-20260924/
  lab-host is reserved for silicon research, so images must build elsewhere. This track made the laptop (and before it build-host-A) able to run the same AVED build_all pipeline as Oleg's VM, and carries the fit 
- page:pair-silicon-c8cf-baseline-20260924 | 2026-09-24 | historical | Pair silicon baseline: HEAD 0xE4B711C8 + TAIL 0xE4B711CF — matched benchmark, width and depth sweeps [0xE4B711C8,0xE4B711C9,0xE4B711CF]
  https://hardware.research.yshicarrick.com/p/pair-silicon-c8cf-baseline-20260924/
  First full characterization of the EAGLE4 HEAD 0xE4B711C8 + TAIL 0xE4B711CF pair image on V80 silicon after the 2026-09-24 cold AC: the matched mt_bench benchmark, a runtime tree-width sweep (1..8) wi
- page:eagle4-restore-g4-embfold-20260924 | 2026-09-24 | historical | G4 embedding constant fold: HBM P-table, fetch engine, host path [0xE4B711C9]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-g4-embfold-20260924/
  <p>G4 (the QKV GEMV) is the largest HEAD phase. Half of its input, the e_norm of the token embedding, depends only on the row. G4 is all-integer until one final rounding, so starting each lane's 48-bi
- page:eagle4-restore-array-poc-20260924 | 2026-09-24 | historical | Five speed levers on the 2-bit GEMV array + adder-tree study (module-only PoC) [0xE4B71155,0xE4B711C9,0xE4B711D3]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-array-poc-20260924/
  The GEMV array that serves G4/G6/G7g/G7u/G9/G12 was taken out of the drafter and five speed levers were tried on it alone: every variant is bit-exact against the unmodified array in xsim, cycle-counte
- page:eagle4-restore-arch-page-20260924 | 2026-09-24 | superseded | Where the EAGLE4 pair's drafter lives on silicon [0xE4B711B0,0xE4B711C9,0xE4B711D2]
  https://hardware.research.yshicarrick.com/p/eagle4-restore-arch-page-20260924/
  A measured map of the EAGLE4 2-bit drafter as it runs on two AMD V80 cards (HEAD 0xE4B711C9 G1–G7u, TAIL 0xE4B711B0 G7p–G8, 102.13 tok/s at accept 2.22): what each phase costs, where each block and me
- page:mesh-queue-dispatch-20260924 | 2026-09-24 | historical | Mesh job queue and dispatch orchestrator
  https://hardware.research.yshicarrick.com/p/mesh-queue-dispatch-20260924/
  Submitters used to pick machines themselves; lab-host had three research jobs waiting 35+ minutes for RAM while two new 251 GB cloud VMs sat idle. The queue separates WHAT to run (submitters) from WHERE (
- page:example-array-poc-20260923 | 2026-09-23 | superseded | Five speed levers on the 2-bit GEMV array (module-only PoC) [0xE4B71155,0xE4B711C9,0xE4B711D3]
  https://hardware.research.yshicarrick.com/p/example-array-poc-20260923/
  The GEMV array that serves G4/G6/G7g/G7u/G9/G12 was taken out of the drafter and five speed levers were tried on it alone: every variant is bit-exact against the unmodified array in xsim, cycle-counte

## skill (41)

- skill:eagle4-aved-build | 2026-09-24 | current | eagle4-aved-build
  <repo>/.claude/skills/eagle4-aved-build/SKILL.md
  Launch and babysit an EAGLE4 AVED PDI build (uni_pipeline) on lab-host / Oleg VM / build-host-A / laptop — Vivado setup, deriving a pipeline, env vars, locks + memory holds, which files to change (eagle4_config
- skill:eagle4-flash-coldac | 2026-09-24 | current | eagle4-flash-coldac [0xE4B711B0,0xE4B711CA]
  <repo>/.claude/skills/eagle4-flash-coldac/SKILL.md
  Flash an EAGLE4 pair image to a V80 (HEAD 61:00.0 / TAIL 41:00.0) and bring the pair back after the user's cold AC — preflash adjudication, detached -p 0 flash, post-cold-AC sequence, PERSIST arming, 
- skill:eagle4-mesh-housekeeping | 2026-10-16 | current | eagle4-mesh-housekeeping
  <repo>/.claude/skills/eagle4-mesh-housekeeping/SKILL.md
  Housekeeping for the EAGLE4 compute mesh — machines/IPs/access, Vivado + license per host, swap/sshd, the mesh dispatcher (mesh_run/status/adopt/install), and the three web sites (mesh RAM monitor, IP
- skill:eagle4-pair-silicon-test | 2026-09-24 | current | eagle4-pair-silicon-test
  <repo>/.claude/skills/eagle4-pair-silicon-test/SKILL.md
  Test an EAGLE4 HEAD/TAIL pair image on silicon after a cold AC — bring-up, arming, the matched benchmark, width/depth sweeps with paired A/Bs, per-phase census, GPU goldens, per-level and per-op corre
- skill:eagle4-restore-arch-page | 2026-09-24 | current | eagle4-restore-arch-page [0xE4B711B0,0xE4B711C9,0xE4B711CA,0xE4B711D2...]
  <repo>/.claude/skills/eagle4-restore-arch-page/SKILL.md
  Resume the interactive architecture page track of the EAGLE4 pair — a published Artifact (EAGLE4 Pair Silicon) mapping phases, GEMV lanes, memory media, HBM traffic, per-die placement, measured GEMM l
- skill:eagle4-restore-array-dualclk | 2026-09-24 | current | eagle4-restore-array-dualclk [0xE4B711C8,0xE4B711C9,0xE4B711CD]
  <repo>/.claude/skills/eagle4-restore-array-dualclk/SKILL.md
  Resume the fast-array-clock track of the EAGLE4 pair. The 2-bit GEMV array (G4/G6/G7g/G7u) runs on a 132 MHz clock that is exactly 2x the drafter's 66 MHz, fed from a per-die BRAM activation cache. It
- skill:eagle4-restore-array-poc | 2026-09-24 | current | eagle4-restore-array-poc [0xE4B711C9]
  <repo>/.claude/skills/eagle4-restore-array-poc/SKILL.md
  Resume the 2-bit GEMV array PoC track of the EAGLE4 pair — levers 1-5 (locality/clock, LANES256/K_PAR32, DSP MACs, TILE_PIPE, G12 shortlist) + merges M1/M2/M3 + the adder-tree study; DONE except M3/DO
- skill:eagle4-restore-engine-split-pair | 2026-09-24 | current | eagle4-restore-engine-split-pair
  <repo>/.claude/skills/eagle4-restore-engine-split-pair/SKILL.md
  Resume the engine-split design-D STEP 3 track of the EAGLE4 pair — GEMV card (role 3) + state card (role 4) + real hop_link in ONE xsim, bit-exact vs the single card (role 0), with RED arms and the jo
- skill:eagle4-restore-fmax-census | 2026-09-24 | current | eagle4-restore-fmax-census [0xE4B711B0,0xE4B711C9,0xE4B711CA,0xE4B711CD]
  <repo>/.claude/skills/eagle4-restore-fmax-census/SKILL.md
  Resume the whole-drafter fmax census track of the EAGLE4 pair — retime the ROUTED full-chip checkpoints of HEAD 0xE4B711C9 / TAIL 0xE4B711CA to 100 and 133 MHz, bucket every failing endpoint by phase 
- skill:eagle4-restore-g4-embfold | 2026-09-24 | current | eagle4-restore-g4-embfold
  <repo>/.claude/skills/eagle4-restore-g4-embfold/SKILL.md
  Resume the G4 embedding-constant-fold (G4_EMBFOLD) track of the EAGLE4 pair HEAD — P[path][token] table in HBM preloads G4's 48-b accumulators (−46.8% G4 cycles, bit-exact). State 2026-09-24: builder 
- skill:eagle4-restore-gpu-accept-grid | 2026-09-24 | current | eagle4-restore-gpu-accept-grid
  <repo>/.claude/skills/eagle4-restore-gpu-accept-grid/SKILL.md
  Resume the GPU-only accept/tok-s grid track of the EAGLE4 pair — H200 SGLang 0.5.16 exec-mode -1, tree width {3,4,6,8} x depth {2,3,4,7,15} x {mt_bench cap-32, humaneval} x {fa3,fa4}, plus batch-size-
- skill:eagle4-restore-head-levers | 2026-09-24 | current | eagle4-restore-head-levers [0xE4B71137,0xE4B711CD]
  <repo>/.claude/skills/eagle4-restore-head-levers/SKILL.md
  Resume the HEAD-levers track of the EAGLE4 pair — SEAM_WIDE (256-bit per-child seam export, modes 0/1/2) + KVONLY_SKIP_ATTN (KV-only rows append K/V and skip score/softmax/PV) in fork head_seam256_kvs
- skill:eagle4-restore-host-overhead | 2026-09-24 | current | eagle4-restore-host-overhead [0xE4B711B0,0xE4B711C9,0xE4B711CA]
  <repo>/.claude/skills/eagle4-restore-host-overhead/SKILL.md
  Resume the pair host round-trip overhead track of the EAGLE4 pair. The track profiles and cuts the SGLang host glue around the HEAD/TAIL pair. HOSTFAST=1 is measured at +4.8% and is now the launcher d
- skill:eagle4-restore-laptop-build-d6 | 2026-09-24 | current | eagle4-restore-laptop-build-d6 [0xE4B711CF,0xE4B711D3,0xE4B711D6]
  <repo>/.claude/skills/eagle4-restore-laptop-build-d6/SKILL.md
  Resume the laptop-build-host + D6 fit-probe track of the EAGLE4 pair — the laptop (Legion-7) and build-host-A as full PDI build hosts, and probe 0xE4B711D6 (GEMV card, attention pruned, LANES 256 + RET_SER)
- skill:eagle4-restore-research-site | 2026-09-24 | current | eagle4-restore-research-site
  <repo>/.claude/skills/eagle4-restore-research-site/SKILL.md
  Resume the research-site track of the EAGLE4 pair — the public page site hardware.research.yshicarrick.com (Cloudflare Worker + D1) where every agent pushes its own experiment page with push_page.py a
- skill:eagle4-restore-sibling-batching | 2026-09-24 | current | eagle4-restore-sibling-batching
  <repo>/.claude/skills/eagle4-restore-sibling-batching/SKILL.md
  Resume the sibling-batching (W columns per weight sweep) track of the EAGLE4 pair — array RTL + gates done and committed, W=4 L128 placed (SLL 821/516), W=8 blocked by AREA not SLL; W=8 L64/L128 place
- skill:eagle4-restore-state-card | 2026-09-24 | current | eagle4-restore-state-card [0xE4B711D0,0xE4B711D1]
  <repo>/.claude/skills/eagle4-restore-state-card/SKILL.md
  Resume the engine-split STATE card track (design D step 2, PIPE_ROLE 4) of the EAGLE4 pair — also folds in step 1 (GEMV card, role 3) and the hop_link transport study. DONE: role-4 RTL xsim-GREEN bit-
- skill:eagle4-restore-systolic-lutgemm | 2026-09-24 | current | eagle4-restore-systolic-lutgemm [0xE4B711C9]
  <repo>/.claude/skills/eagle4-restore-systolic-lutgemm/SKILL.md
  Resume the systolic-array + multiply-free LUT-GEMM track of the EAGLE4 pair — two bit-exact variants of the 2-bit GEMV array (A = systolic activation chain, B = shared bit-plane table lane). State: al
- skill:eagle4-restore-tree-width | 2026-09-24 | current | eagle4-restore-tree-width [0xE4B711C9,0xE4B711CB,0xE4B711CE,0xE4B711CF]
  <repo>/.claude/skills/eagle4-restore-tree-width/SKILL.md
  Resume the runtime tree-width track of the EAGLE4 pair — AXI-lite 0x1CC tree_width (1..8, reset 8) on the TAIL so the pair runs width 3 (modelled +27–35% at d2); TAIL image CF (0xE4B711CF = fast TAIL 
- skill:eagle4-timing-debug | 2026-10-17 | current | eagle4-timing-debug
  <repo>/.claude/skills/eagle4-timing-debug/SKILL.md
  Debug a failed V80/AVED PDI timing closure end to end. Read route progress by artifact; triage the routed checkpoint by clock pair and instance; re-time with constraint-only fixes to decide whether th
- skill:dispatch-agent | 2026-08-08 | current | dispatch-agent
  <repo>/.claude/skills/dispatch-agent.md
  Launch a background agent the EAGLE4 way — correct model tier, standing safety rules, resource envelope, and deliverable contract baked into every brief.
- skill:dma-driver | 2026-05-23 | current | dma-driver
  <repo>/.claude/skills/dma-driver.md
  Write or modify a QDMA host-side DMA / register-transfer driver for the V80, or wire the FPGA drafter into SGLang's GPU loop. Use when touching host transfer code, the AxiLiteRegs harness, the s_axili
- skill:eagle4-doomsday | 2026-09-11 | current | eagle4-doomsday [0xE4B71110,0xE4B71118,0xE4B71127,0xE4B71128...]
  <repo>/.claude/skills/eagle4-doomsday.md
  Full operating capacity for the EAGLE4 FPGA project, written for an agent with ZERO prior context. The whole pipeline end to end with exact verified commands (GPU golden -> xsim -> OOC synth -> build_
- skill:feedback-loop | 2026-08-08 | historical | feedback-loop
  <repo>/.claude/skills/feedback-loop.md
  The EAGLE4 rule-promotion procedure — how corrections become binding, and the standing operating loop (13 rules) to re-read on resume.
- skill:flash-test-verify | 2026-04-12 | historical | flash-test-verify
  <repo>/.claude/skills/flash-test-verify.md
  Flash PDI to V80, run golden gate tests, benchmark SGLang E2E, compare against GPU baseline
- skill:hls-design | 2026-04-26 | historical | hls-design
  <repo>/.claude/skills/hls-design.md
  HLS kernel design for EAGLE4 V80 — pick primitives, copy-and-prune from v56, apply pragma cheatsheet, avoid known traps. Use when starting a new block / modifying existing kernel source.
- skill:hls-experiment | 2026-04-13 | historical | hls-experiment
  <repo>/.claude/skills/hls-experiment.md
  Run a full HLS experiment cycle: create variant source, C-sim verify (3 golden gates), launch HLS synthesis, extract csynth metrics, update tracker
- skill:hls-synth | 2026-04-26 | historical | hls-synth
  <repo>/.claude/skills/hls-synth.md
  Run Vitis HLS synthesis for an EAGLE4 kernel, parse csynth.rpt for resources/slack/Fmax, decide PDI-ready or iterate. Use after kernel C-sim passes.
- skill:hls-visual-map | 2026-04-27 | historical | hls-visual-map
  <repo>/.claude/skills/hls-visual-map.md
  Maintain visual, hardware-aware documentation of HLS code under docs/hls-visual-map/. Update editable diagram sources + SVG renders + PDF + manifest + changelog every time meaningful HLS code changes.
- skill:monitor-builds | 2026-08-08 | historical | monitor-builds
  <repo>/.claude/skills/monitor-builds.md
  Arm watchers and monitors for builds/sims the EAGLE4 way — notification-driven, never polling in the orchestrator context; verify compute not logs.
- skill:pdi-build | 2026-04-26 | historical | pdi-build
  <repo>/.claude/skills/pdi-build.md
  Build a flashable PDI from an HLS-synthesized kernel via AVED build_all.sh — install IP into iprepo, set VLNV env, monitor synth/opt/place/route/bitgen phases, capture post-route WNS. Use after /hls-s
- skill:pdi-doomsday-sprint | 2026-08-10 | historical | pdi-doomsday-sprint
  <repo>/.claude/skills/pdi-doomsday-sprint.md
  
- skill:pdi-flash | 2026-04-26 | historical | pdi-flash
  <repo>/.claude/skills/pdi-flash.md
  Flash a PDI to V80, coordinate cold AC reboot with user, set up QDMA queues, run silicon validation. Use after /pdi-build with positive post-route WNS.
- skill:vitis-hls-synthesis | 2026-04-12 | historical | vitis-hls-synthesis
  <repo>/.claude/skills/vitis-hls-synthesis.md
  ---
- skill:vivado-analysis | 2026-04-12 | historical | vivado-analysis
  <repo>/.claude/skills/vivado-analysis.md
  ---
- skill:vivado-constraints | 2026-04-12 | historical | vivado-constraints
  <repo>/.claude/skills/vivado-constraints.md
  ---
- skill:vivado-debug | 2026-04-12 | historical | vivado-debug
  <repo>/.claude/skills/vivado-debug.md
  ---
- skill:vivado-impl | 2026-04-12 | historical | vivado-impl
  <repo>/.claude/skills/vivado-impl.md
  ---
- skill:vivado-sim | 2026-04-12 | historical | vivado-sim
  <repo>/.claude/skills/vivado-sim.md
  ---
- skill:vivado-synth | 2026-04-12 | historical | vivado-synth
  <repo>/.claude/skills/vivado-synth.md
  ---
- skill:vivado-tcl | 2026-04-12 | historical | vivado-tcl
  <repo>/.claude/skills/vivado-tcl.md
  ---

## handoff (13)

- handoff:00_START_HERE.md | 2026-09-24 | current | HANDOFF 2026-09-24 — START HERE (orchestrator session 87e06065, 2026-09-23 ~09:00 → 2026-09-24 ~02:3 [0xE4B711B0,0xE4B711C8,0xE4B711C9,0xE4B711CA...]
  <repo>/docs/handoff_20260924/00_START_HERE.md
  Written by the orchestrator at the user's request ("record all you learned so far … so a fresh you with no
- handoff:01_PROJECT_AND_ARCHITECTURE.md | 2026-09-24 | current | 01 — Project and architecture (what the thing is, how time is spent) [0xE4B711C9]
  <repo>/docs/handoff_20260924/01_PROJECT_AND_ARCHITECTURE.md
  EAGLE4 speculative decoding for Llama-3.2-3B (target, on an H200 GPU) with a 2-bit EAGLE4 drafter running on AMD
- handoff:02_2BIT_ENGINE_RESEARCH.md | 2026-09-24 | current | 02 — The 2-bit GEMV engine: everything measured tonight
  <repo>/docs/handoff_20260924/02_2BIT_ENGINE_RESEARCH.md
  The user asked "how do we make the 2-bit array wayyy faster". Only width (LANES, K_PAR) and clock can help: the array
- handoff:03_ENGINE_SPLIT_DESIGN_D.md | 2026-09-24 | current | 03 — Engine split, design D (the chosen next architecture) [0xE4B711D3]
  <repo>/docs/handoff_20260924/03_ENGINE_SPLIT_DESIGN_D.md
  Decision doc: `reports/results/TWO_CARD_ARCHITECTURE_DECISION_20260923.md`; design doc
- handoff:04_TAIL_TREE_WIDTH_BATCHING.md | 2026-09-24 | current | 04 — TAIL speed, runtime tree width, sibling batching, the G4 embedding fold [0xE4B711CF]
  <repo>/docs/handoff_20260924/04_TAIL_TREE_WIDTH_BATCHING.md
  TISEQ=1 + TREE_P=16 + SER_DEPTH=16384 on top of CA's source. Full-chain TAIL xsim, 18 real OSPREY items, old vs new
- handoff:05_MODELS_ACCEPT_AND_WIN_CONDITION.md | 2026-09-24 | current | 05 — Speed models, the win condition vs GPU, the accept gap
  <repo>/docs/handoff_20260924/05_MODELS_ACCEPT_AND_WIN_CONDITION.md
  These answer "can the FPGA pair beat the GPU, and what matters most?" Read them before proposing any new lever.
- handoff:06_IMAGES_BUILDS_FLASH_TEST.md | 2026-09-24 | current | 06 — Images, builds, flash, cold AC, silicon tests [0xE4B711B0,0xE4B711C8,0xE4B711C9,0xE4B711CA...]
  <repo>/docs/handoff_20260924/06_IMAGES_BUILDS_FLASH_TEST.md
  Build-ids are `0xE4B711xx`; `PROBE_REV 8'hxx` is the low byte. Always grep scripts/, reports/results/ and NOW.md for zero
- handoff:07_INFRA_MACHINES_MESH_SITES.md | 2026-09-24 | current | 07 — Machines, access, licenses, the mesh dispatcher, the web sites
  <repo>/docs/handoff_20260924/07_INFRA_MACHINES_MESH_SITES.md
  Setup facts: build-host-A is native Ubuntu 24.04 (the user calls it "the PC"); Vivado at `/shared/tools/xilinxinstall/2025.1`
- handoff:08_AGENTS_OPEN_ITEMS_NEXT.md | 2026-09-24 | current | 08 — Agent roster, running jobs, decisions waiting on the user, next steps
  <repo>/docs/handoff_20260924/08_AGENTS_OPEN_ITEMS_NEXT.md
  Agents are sub-sessions of the orchestrator; they die with the session (and with a lab-host cold AC). Their durable output
- handoff:09_LESSONS_AND_TRAPS.md | 2026-09-24 | current | 09 — Lessons and traps from this session (each one cost time tonight)
  <repo>/docs/handoff_20260924/09_LESSONS_AND_TRAPS.md
  1. "Locality lifts the array to ~167 MHz" — the control run showed the array ALONE closes ~166 MHz with either
- handoff:10_AGENT_RESTORE_SKILLS.md | 2026-09-24 | current | 10 — Agent restore skills: each agent writes its own, at an assigned path [0xE4B711CD,0xE4B711CF]
  <repo>/docs/handoff_20260924/10_AGENT_RESTORE_SKILLS.md
  User directive 2026-09-24 ~02:45: "and all the agent to restore: let the agent's themselves write these, and put to
- handoff:11_BRINGUP_TEST_PLAN_C8_CF.md | 2026-09-24 | current | 11 — Bring-up + test plan after the cold AC that boots C8 (HEAD) + CF (TAIL)  (written 2026-09-24 11 [0xE4B711B0,0xE4B711C8,0xE4B711C9,0xE4B711CD...]
  <repo>/docs/handoff_20260924/11_BRINGUP_TEST_PLAN_C8_CF.md
  The lab-host cold AC kills this session and every agent. The next session starts here + skill `eagle4-flash-coldac`.
- handoff:12_AGENT_FLEET_SPAWN_LIST.md | 2026-09-24 | current | 12 — Agent fleet spawn list (session 87e06065, 2026-09-23 09:00 → 2026-09-24 11:40) [0xE4B711CD,0xE4B711CF,0xE4B711E1,0xE4B711E3]
  <repo>/docs/handoff_20260924/12_AGENT_FLEET_SPAWN_LIST.md
  How the orchestrator launched and ran its fleet, and every agent it spawned — so a fresh session can re-spawn the same

## repo_key_doc (10)

- repo:NOW.md | 2026-10-18 | current | NOW — EAGLE4 CURRENT STATE  ·  last updated 2026-09-14 06:00 EDT (**CARD/PDI STATE UNCHANGED**: both [0xE4B7111B,0xE4B71126,0xE4B71127,0xE4B71132...]
  <repo>/NOW.md
  - **Tracks (detect by ARTIFACT):** A BD tcl `scripts/qsfp_seam_head/qsfp_seam_bd.tcl` + `QSFP_SEAM_BD_INTEGRATION_20260911.md` · B golden TB `tb/tb_qsfp_seam_adapter.sv` + `QSFP_SEAM_ADAPTER_TB_202609
- repo:TRACKLIST_20260923.md | 2026-09-23 | current | TRACKLIST — architecture-forward board (started 2026-09-23) [0xE4B7118B,0xE4B71192,0xE4B71193,0xE4B7119C...]
  <repo>/TRACKLIST_20260923.md
  Editable. **Update this file in the same turn a job starts, lands or changes state.** Status: `RUNNING` · `QUEUED` · `PAUSED` · `DONE` · `DECIDE` (waiting on the user).
- repo:USER_DIALOGUE_LOG.md | 2026-10-23 | current | USER DIALOGUE LOG — read every resume/tick; the cron + compaction make me forget, this is the anchor [0xE4B71107,0xE4B71109,0xE4B7110C,0xE4B71117...]
  <repo>/USER_DIALOGUE_LOG.md
  **Purpose (user-mandated 2026-06-16):** the cron/compaction loop drops user intent. This file is the durable
- repo:INDEX.md | 2026-07-13 | current | EAGLE4 FPGA Project — Master Index (markdown knowledge + code locations) [0xE4B71109,0xE4B7110A,0xE4B7110B,0xE4B7110C...]
  <repo>/INDEX.md
  **Project root:** `./`
- repo:CLAUDE.md | 2026-10-22 | current | EAGLE4 FPGA Project — Claude Code Entry Point [0xE4B71118,0xE4B71127,0xE4B71143,0xE4B71155...]
  <repo>/CLAUDE.md
  **You are working on accelerating the EAGLE4 speculative-decoding drafter on AMD Alveo V80 (Versal HBM, AVED 25.1, QDMA driver — NOT XRT). Target: Llama-3.2-3B + 2-bit EAGLE4 drafter.**
- repo:README.md | 2026-09-27 | current | EAGLE4 pair — tested archive (lab-host, snapshot 2026-09-27 ~02:10) [0xE4B711A5,0xE4B711F4]
  <repo>/tested_archive/README.md
  This folder is the reproducibility archive for every HEAD/TAIL V80 pair image tested on lab-host in the 2026-09-21..27 silicon campaign.
- repo:GPU_ACCEPT_GOLDEN_COMMAND.md | 2026-08-25 | current | 📌 PINNED — THE GPU ACCEPT REFERENCE COMMAND (user-supplied 2026-08-24 06:01)
  <repo>/docs/GPU_ACCEPT_GOLDEN_COMMAND.md
  **This is the command. Do not reconstruct it, do not substitute paths, do not "improve" it.**
- repo:RESEARCH_SITE.md | 2026-10-22 | current | Research site: hardware.research.yshicarrick.com
  <repo>/docs/RESEARCH_SITE.md
  One page per research experiment, written and pushed by the agent that ran it. Public read.
- repo:PROJECT_TIMELINE_AND_INVARIANTS.md | 2026-07-06 | current | EAGLE4 V80 Drafter — Project Timeline & Must-Not-Forget Invariants
  <repo>/docs/PROJECT_TIMELINE_AND_INVARIANTS.md
  **Purpose:** survive context compaction. Every session should read this first. It holds (A) the
- repo:STATIC_HUNTER_CHARTER.md | 2026-06-19 | current | STATIC HUNTER CHARTER — read this first on every (re)launch
  <repo>/docs/STATIC_HUNTER_CHARTER.md
  You are a **long-lived static anomaly hunter** for the EAGLE4 FPGA drafter. Your job is ONE thing:

## repo_report (132)

- report:AGENT_WORK_TRACKER.md | 2026-07-29 | historical | AGENT WORK TRACKER — all agent-driven work, artifacts, data, and source modifications [0xE4B71112,0xE4B71113]
  <repo>/reports/AGENT_WORK_TRACKER.md
  1. **A5 lands** → preflash dry-run + report; flash/cold-AC = USER decision. Then J5-α silicon: occupancy re-test with real backend.
- report:ANALYSIS_STATE_AND_NEXT_PDI_20260713.md | 2026-07-13 | historical | ANALYSIS — Current Numbers, Directions, Proofs & Validations → Next PDI Decision [0xE4B7110F,0xE4B71110,0xE4B71111,0xE4B71112]
  <repo>/reports/ANALYSIS_STATE_AND_NEXT_PDI_20260713.md
  RTL change (own fork, verilog/forks/<name>/)
- report:ARCH_ONETIME_PACKAGE.md | 2026-07-19 | historical | ONE-TIME ARCHITECTURE PACKAGE — consolidated findings + the real integration ladder
  <repo>/reports/ARCH_ONETIME_PACKAGE.md
  Baseline: 3,167 µs @66 MHz (210,118 cyc) silicon.
- report:ARTIFACT_POINTERS.md | 2026-09-14 | current | ARTIFACT_POINTERS.md — netlist + cycle-evidence paths (D167, 2026-08-03T16:52:22-04:00; every row ve [0xE4B71129,0xE4B7113F,0xE4B71142]
  <repo>/reports/ARTIFACT_POINTERS.md
  - [OK] `verilog/forks/w_dispatch/census_sbx/serialtail/SHIPCFG/xsim.log` — schedule-MEASURED child-pass census (the ±0.4% instrument's ground truth)
- report:AUTONOMOUS_DRIVE_LOG.md | 2026-08-26 | historical | Autonomous Drive Log — tree → PDI on silicon [0xE4B71111,0xE4B71115,0xE4B71118,0xE4B7111E...]
  <repo>/reports/AUTONOMOUS_DRIVE_LOG.md
  **Purpose:** bootstrap record so any resumed/compacted instance can continue without re-deriving.
- report:BEST_SPEED_INTEGRATION.md | 2026-06-14 | historical | BEST-SPEED INTEGRATION — the consolidated additional-PDI top (build #2-speed)
  <repo>/reports/BEST_SPEED_INTEGRATION.md
  **Date:** 2026-06-14
- report:BITNET_MIGRATION_PLAN.md | 2026-07-21 | historical | BITNET MIGRATION PLAN — 2-bit GPTQ → BitNet ternary (Phase 2 of BITNET_MIGRATION_TRACKLIST.md)
  <repo>/reports/BITNET_MIGRATION_PLAN.md
  1. **The migration is much smaller than the tensor-shape scare suggested.** The shipped RTL is
- report:BITNET_MIGRATION_TRACKLIST.md | 2026-07-21 | historical | BITNET MIGRATION — TRACKLIST (step-by-step, no rushing; every step leaves an artifact)
  <repo>/reports/BITNET_MIGRATION_TRACKLIST.md
  Migrate the FPGA drafter from 2-bit GPTQ (accept 2.08) to the BitNet ternary checkpoint (author-reported
- report:BITNET_PARKED_STATE.md | 2026-07-21 | historical | BITNET MIGRATION — PARKED STATE (the single entry point for resuming) [0xE4B71115,0xE4B71117]
  <repo>/reports/BITNET_PARKED_STATE.md
  **User directive, verbatim, 2026-07-21** (`USER_DIALOGUE_LOG.md:278-279`):
- report:BLOCK4_BITEXACT_EVIDENCE_AUDIT_20260518.md | 2026-05-18 | historical | Block 4 Bit-Exact Evidence Audit — Forensic Reconstruction
  <repo>/reports/BLOCK4_BITEXACT_EVIDENCE_AUDIT_20260518.md
  **Date:** 2026-05-18 → 19 (overnight sprint)
- report:BLOCK4_HEADLINE_RETRACTION_NEEDED.md | 2026-06-05 | historical | Block 4 Headline Retraction — Editing Surface
  <repo>/reports/BLOCK4_HEADLINE_RETRACTION_NEEDED.md
  **Date raised:** 2026-05-20
- report:BLOCK4_LATENCY_BREAKDOWN_20260520.md | 2026-05-20 | historical | Block 4 (dummy_layer) Latency Breakdown — 2026-05-20 spatial-fix synth
  <repo>/reports/BLOCK4_LATENCY_BREAKDOWN_20260520.md
  **Source:** `EAGLE/synthesis_bundle/sparsity_experiments/exp11_monolithic/hls_dummy_layer/eagle4_drafter_dummy_layer/hls/syn/report/csynth.rpt`. Header line 7: Wed May 20 03:22:18 2026, Vitis 2025.1 b
- report:BODY_SHRINK_RANKING.md | 2026-07-31 | historical | BODY_SHRINK_RANKING — Lever G: the post-gather body floor & ranked accuracy-neutral body-shrink leve
  <repo>/reports/BODY_SHRINK_RANKING.md
  **Date:** 2026-06-13
- report:BUILD_LADDER_AND_FIXES.md | 2026-08-19 | historical | BUILD LADDER & FIXES — tree-FSM PDI campaign (2026-07-25 → 07-26) [0xE4B71118]
  <repo>/reports/BUILD_LADDER_AND_FIXES.md
  Flash remains **user + physical cold-AC gated**; no image flashes on negative WNS; only final FPT PDIs
- report:CONFIG_PARAMETER_REFERENCE.md | 2026-07-31 | historical | CONFIG PARAMETER REFERENCE [0xE4B71111,0xE4B71120,0xE4B71121,0xE4B71125...]
  <repo>/reports/CONFIG_PARAMETER_REFERENCE.md
  **Date** 2026-07-27 (values re-schemed 2026-07-31) · **Primary fork** `verilog/forks/sw_treebatch_w2` (config header md5 `41c2c1362c5262f7efa0529ceca93428`, 76 straps — ⚠ **stale, see above: live md5 
- report:CONSOLIDATED_KNOWLEDGE.md | 2026-09-02 | current | EAGLE4 — CONSOLIDATED KNOWLEDGE (single-file rollup of 1,330 markdowns) [0xE4B71101,0xE4B71143]
  <repo>/reports/CONSOLIDATED_KNOWLEDGE.md
  **Built 2026-06-22.** This file consolidates the project's markdown sprawl into one durable rollup.
- report:CONTEXT_RECOVERY_TAKEOVER.md | 2026-07-23 | historical | CONTEXT RECOVERY / AGENTIC TAKEOVER — state as of 2026-07-23 ~13:30 EDT [0xE4B71115,0xE4B71118,0xE4B71119,0xE4B7111B]
  <repo>/reports/CONTEXT_RECOVERY_TAKEOVER.md
  **In-flight processes right now** (verify with `ps -eo etime,comm | grep -E 'xsimk|vivado'`):
- report:CURRENT_WORKING_IMAGE_BASELINE.md | 2026-06-13 | historical | Current Working PDI Image + Testing Method — BASELINE (confirmed 2026-06-13) [0xE4B71101]
  <repo>/reports/CURRENT_WORKING_IMAGE_BASELINE.md
  This is the canonical "what's on the card right now and how to test it" reference. Re-confirm
- report:DECISION_LOG.md | 2026-09-06 | current | DECISION LOG — EAGLE4 (append-only, newest first) [0xE4B71118,0xE4B71143,0xE4B71146,0xE4B71159...]
  <repo>/reports/DECISION_LOG.md
  Split out of `NOW.md` on 2026-08-24 (D795). NOW.md is the CURRENT-STATE index and
- report:DEPTHN_EMB_GATHER_WORKORDER.md | 2026-07-24 | historical | DEPTH-N EMBEDDING GATHER — findings + build-host-M work order (merge-bar: xsim GREEN + OOC synth clean)
  <repo>/reports/DEPTHN_EMB_GATHER_WORKORDER.md
  **Net (re-corrected 2026-07-24):** depth-N needs (1) an on-chip embedding fetch engine, (2) the `emb_buf` write port, AND (3) a **new HBM residency for the embedding table** (~751.5 MB @bf16, or ~192 
- report:DEPTHN_GAP_AND_PLAN.md | 2026-06-13 | historical | Depth-N Self-KV Drafter — Gap to Replacing the SGLang GPU Drafter (2026-06-13)
  <repo>/reports/DEPTHN_GAP_AND_PLAN.md
  Bit-exact oracle = `reference/src/chain.cpp::drafter_tree_decode` (== GPU `select_top_k_tokens`).
- report:DEPTHN_GOLDEN_REFERENCE.md | 2026-07-24 | historical | Depth-N Golden Reference
  <repo>/reports/DEPTHN_GOLDEN_REFERENCE.md
  **Created** 2026-07-21 · **Status** captures GREEN, reference model GREEN, RED arms 5/5 detected
- report:DEPTHN_INPLACE_DESIGN.md | 2026-06-13 | historical | In-Place Depth-N Tree Decoding — RTL Integration Design (2026-06-13)
  <repo>/reports/DEPTHN_INPLACE_DESIGN.md
  **Goal:** make the FPGA the *complete* EAGLE4 drafter — expand the candidate tree to depth N
- report:DEPTHN_build-host-M_MIGRATION.md | 2026-07-22 | historical | Depth-N RUNG-C Correctness Sim — lab-host → build-host-M Migration [0xE4B71119]
  <repo>/reports/DEPTHN_build-host-M_MIGRATION.md
  **What the lab-host run had reached (11 h in):** live-G8 real-body path CONFIRMED
- report:DEPTHN_RUNGC_SIM_DOSSIER.md | 2026-07-24 | historical | DEPTH-N / RUNG-C xsim DOSSIER — the "11-hour sim" (`rungc_depthn2`), fully quoted [0xE4B71115,0xE4B71118,0xE4B71119]
  <repo>/reports/DEPTHN_RUNGC_SIM_DOSSIER.md
  **The north star.** Depth-N on-chip tree decode is the user's headline goal since 2026-06-09
- report:DRAFTER_MODEL_HW_TRANSLATION.md | 2026-07-31 | historical | DRAFTER-MODEL ↔ HARDWARE TRANSLATION — meeting doc for the drafter-training team [0xE4B71111]
  <repo>/reports/DRAFTER_MODEL_HW_TRANSLATION.md
  Forward cycle split [XSIM, TB-matched to silicon <1%]: **body 80,093 (38%) + tree/G12 walk 28,104 (14%) + G8 gather 101,921 (49%)**; gather = 512 rows × 196.27 cyc/row [SILICON-matched].
- report:EVIDENCE_BODY_SYSTEM.md | 2026-09-02 | current | EVIDENCE_BODY_SYSTEM — Consolidated claim→data→source→tier ledger
  <repo>/reports/EVIDENCE_BODY_SYSTEM.md
  **Date:** 2026-06-14
- report:EVIDENCE_CORRECTNESS.md | 2026-06-14 | historical | EVIDENCE-TIERED CORRECTNESS LEDGER — pdi11_grand (night of 2026-06-13)
  <repo>/reports/EVIDENCE_CORRECTNESS.md
  Read-only consolidation. Every row is `claim → test → measured data → file citation → tier`.
- report:EVIDENCE_GATHER_LEVERS.md | 2026-07-31 | historical | EVIDENCE-TIERED GATHER-LEVER LEDGER — G8 lm-head rescore shrink/parallelize sweep
  <repo>/reports/EVIDENCE_GATHER_LEVERS.md
  **Date:** 2026-06-14
- report:EVIDENCE_INDUSTRY_ARTIFACTS.md | 2026-06-14 | historical | Evidence Ledger — Industry Surveys + Built Artifacts (2026-06-14) [0xE4B71101]
  <repo>/reports/EVIDENCE_INDUSTRY_ARTIFACTS.md
  **Scope:** READ-ONLY consolidation of tonight's industry surveys and built/tested artifacts.
- report:EXP_G12_BLOCK_PREFETCH_COVERAGE.md | 2026-06-05 | historical | EXP — G12 Block-wise Speculative Prefetch Coverage (Phase A)
  <repo>/reports/EXP_G12_BLOCK_PREFETCH_COVERAGE.md
  **Date:** 2026-06-04
- report:FRESH_AGENT_BRIEFING_VERILOG_SYSTOLIC.md | 2026-06-05 | historical | Briefing for a Fresh Architecture-Exploration Agent
  <repo>/reports/FRESH_AGENT_BRIEFING_VERILOG_SYSTOLIC.md
  **Purpose of this document:** transfer 12 months of project context to an agent that has never seen this codebase, asking them to evaluate a **pure-Verilog + hardwired-2-bit-weights + FP16-systolic-ar
- report:FRSPEC_OOC_CONFIRM.md | 2026-06-14 | historical | FR-Spec URAM-Donor — OOC SYNTHESIS CONFIRMATION
  <repo>/reports/FRSPEC_OOC_CONFIRM.md
  **Date:** 2026-06-14
- report:G7P_KEEP_FIX_RESULT.md | 2026-06-15 | historical | G7P-KEEP Fix — Root-Cause Structural Fix for the Silicon-Ineffective RD_LAT Read Register
  <repo>/reports/G7P_KEEP_FIX_RESULT.md
  **Date:** 2026-06-15
- report:G7P_SILICON_FIX.md | 2026-06-15 | historical | G7P Silicon Fix — Verdict: FIX IS CORRECT AND PRESENT; the "silicon-ineffective" premise was a build [0xE4B71101]
  <repo>/reports/G7P_SILICON_FIX.md
  **Date:** 2026-06-15
- report:G8_NCHAN_STRIPED_GATHER_SPEC.md | 2026-09-02 | current | G8 N-Channel Striped HBM Gather — Design Spec (build-#3 G8 redesign)
  <repo>/reports/G8_NCHAN_STRIPED_GATHER_SPEC.md
  **Status:** ANALYSIS/SPEC ONLY — no build, no xsim, no RTL edits.
- report:GATHER_LEVER_A_KC_REDUCTION.md | 2026-07-31 | historical | Gather Lever A — Smaller K_C (cheap-path pre-prune of the G8 rescore set)
  <repo>/reports/GATHER_LEVER_A_KC_REDUCTION.md
  **Date:** 2026-06-13
- report:GATHER_LEVER_B_PRECISION_RESCORE.md | 2026-06-13 | historical | Gather Lever B — Lower-Precision lm-head Rescore (G8)
  <repo>/reports/GATHER_LEVER_B_PRECISION_RESCORE.md
  **Date:** 2026-06-13  **Mode:** analysis/study only (no build/flash). numpy + `eagle4_ref` (`reference/build/eagle4_ref.cpython-313-x86_64-linux-gnu.so`) + real fp16 lm-head (`reference/model_weights/
- report:GATHER_LEVER_C_PREFETCH_ENABLE.md | 2026-06-13 | historical | GATHER LEVER C — Speculative Prefetch (P2) Enable Analysis
  <repo>/reports/GATHER_LEVER_C_PREFETCH_ENABLE.md
  **Date:** 2026-06-13
- report:GATHER_LEVER_D_ROW_CACHE.md | 2026-06-13 | historical | Gather Lever D — Cross-Forward lm-head ROW CACHE
  <repo>/reports/GATHER_LEVER_D_ROW_CACHE.md
  **Date:** 2026-06-13
- report:GATHER_LEVER_E_OUTSTANDING_AR.md | 2026-06-13 | historical | Gather Lever E — DEEPER OUTSTANDING-AR (G8_OUTSTANDING 8 → 16/32/64)
  <repo>/reports/GATHER_LEVER_E_OUTSTANDING_AR.md
  **Date:** 2026-06-13
- report:GATHER_LEVER_F_OVERLAP.md | 2026-07-31 | historical | GATHER LEVER F — Inter-Forward Overlap (G8 this-fwd ‖ body next-fwd)
  <repo>/reports/GATHER_LEVER_F_OVERLAP.md
  **Date:** 2026-06-13
- report:GATHER_LEVER_H_DRAFTVERIFY_OVERLAP.md | 2026-06-13 | historical | GATHER LEVER H — System-Level Draft/Verify Overlap (FPGA-draft(N) ‖ GPU-verify(N-1))
  <repo>/reports/GATHER_LEVER_H_DRAFTVERIFY_OVERLAP.md
  **Date:** 2026-06-13
- report:GATHER_LEVER_I_DOT_II1.md | 2026-06-14 | historical | Gather Lever I — BREAK THE G8 DOT CONSUME FLOOR (II=2 → II=1)
  <repo>/reports/GATHER_LEVER_I_DOT_II1.md
  **Date:** 2026-06-14
- report:GATHER_LEVER_J_CHEAP_FIDELITY.md | 2026-06-13 | historical | Gather Lever J — Cheap-Path Fidelity (can a higher-precision cheap rank resurrect Lever A?)
  <repo>/reports/GATHER_LEVER_J_CHEAP_FIDELITY.md
  **Date:** 2026-06-13
- report:GATHER_LEVER_K_FRSPEC.md | 2026-06-14 | historical | GATHER LEVER K — FR-Spec vocab shortlist on the CHEAP G12 path
  <repo>/reports/GATHER_LEVER_K_FRSPEC.md
  **Date:** 2026-06-14
- report:GATHER_SHRINK_MASTER_RANKING.md | 2026-07-31 | historical | Gather-Shrink Master Ranking — full A–J sweep + WIDE + N-channel (2026-06-14)
  <repo>/reports/GATHER_SHRINK_MASTER_RANKING.md
  The G8 lm-head gather is 93% of the forward (silicon-clock-linear ⇒ **consume-bound**, not
- report:GEMV_COL_DOUBLE_BUFFER.md | 2026-06-05 | historical | GEMV_COL_DOUBLE_BUFFER — column tail-hiding for the scaled SLR GEMV array
  <repo>/reports/GEMV_COL_DOUBLE_BUFFER.md
  **Date:** 2026-06-03
- report:GRAND_M7_ZEROLOGIT_RCA.md | 2026-06-15 | historical | GRAND M7 Zero-Logit RCA [0xE4B71101]
  <repo>/reports/GRAND_M7_ZEROLOGIT_RCA.md
  Symptom: grand M7 = G1 golden, but G4q..G12 = exactly 0x00000000.
- report:INDUSTRY_SOTA_SURVEY.md | 2026-06-14 | historical | Industry & Research SOTA Survey — Memory-Bound LLM Decode, On-Chip Weights, and the 2-bit Spec-Decod
  <repo>/reports/INDUSTRY_SOTA_SURVEY.md
  **Author:** Grand Master PL (FPGA strategy)
- report:INDUSTRY_SPECDECODE_SURVEY.md | 2026-06-14 | historical | Industry Survey: Memory-Bound Decode & Speculative-Decoding Drafters in Production (2024–2026)
  <repo>/reports/INDUSTRY_SPECDECODE_SURVEY.md
  **Date:** 2026-06-14
- report:J5A_HANDOFF_FOR_OPUS.md | 2026-07-15 | historical | J5-α HANDOFF — what Opus must do to carry the full-drafter build forward [0xE4B71111,0xE4B71112,0xE4B71113]
  <repo>/reports/J5A_HANDOFF_FOR_OPUS.md
  - Look for a dated appended section in `reports/results/J5A_ASSEMBLY.md` covering both fixes.
- report:JEWELRY_PDI_MASTER.md | 2026-07-31 | historical | JEWELRY PDI — Master Architecture, Experiment Ledger & Build Plan [0xE4B7110F,0xE4B71111,0xE4B71112]
  <repo>/reports/JEWELRY_PDI_MASTER.md
  **Card: SPEEDWAVE 0xE4B71111 — SILICON-VALIDATED 2026-07-14**
- report:JEWEL_FAILURE_DIAGNOSIS_AND_MITIGATION.md | 2026-07-14 | historical | Jewel J3 Place Failure — Diagnosis, Mitigation Plan, and Implemented Fix [0xE4B7110F]
  <repo>/reports/JEWEL_FAILURE_DIAGNOSIS_AND_MITIGATION.md
  **Log:** `verilog/forks/ps_jewel/build/jewel_slr50/jslr.log` (v1 run, exited Jul 11 06:40)
- report:JEWEL_PHASE_TIME_RESOURCE_DATAFLOW.md | 2026-07-16 | historical | JEWEL LINE — Per-Phase Cycle · Resource · Dataflow Ledger [0xE4B71111,0xE4B71112,0xE4B71113]
  <repo>/reports/JEWEL_PHASE_TIME_RESOURCE_DATAFLOW.md
  Source: `reports/JEWELRY_PDI_MASTER.md` §2, from `reports/results/SPEEDWAVE_PDI_ASSEMBLY.md:69`:
- report:LATENCY_RIGOR_AGENDA.md | 2026-06-05 | historical | Latency Rigor Agenda (2026-05-20 ~05:25 EDT)
  <repo>/reports/LATENCY_RIGOR_AGENDA.md
  The forcing function. Project lead flagged that we are "hand-wavy on latency." This doc collapses three deliverables into a 1-page next-moves list with explicit gaps.
- report:M8_FPGA_SGLANG_INTEGRATION_DESIGN.md | 2026-06-13 | historical | M8 — FPGA-in-the-loop SGLang Integration + Demo (2026-06-14) [0xE4B71101]
  <repo>/reports/M8_FPGA_SGLANG_INTEGRATION_DESIGN.md
  Goal: run the **currently flashed maxprobe image** as the EAGLE4 draft model inside SGLang's
- report:M8_FPGA_SHIM_BUILD.md | 2026-06-13 | historical | M8 FPGA-in-the-loop Shim — Build Report (2026-06-13) [0xE4B71101]
  <repo>/reports/M8_FPGA_SHIM_BUILD.md
  Host-code-only build of the M8 FPGA-in-the-loop SGLang shim + demo per
- report:M8_SGLANG_E2E_RUNBOOK.md | 2026-06-12 | historical | M8 — SGLang FPGA-in-the-Loop E2E Runbook (turnkey for the instant M7 passes)
  <repo>/reports/M8_SGLANG_E2E_RUNBOOK.md
  **Date:** 2026-06-12  **Scope:** host-only prep + runbook. NO flash, NO build, NO card access here.
- report:MAXPROBE_M7_BASELINE_PIN.md | 2026-06-14 | historical | MAXPROBE M7 BASELINE — PROVENANCE PIN (2026-06-14) [0xE4B71101]
  <repo>/reports/MAXPROBE_M7_BASELINE_PIN.md
  Pins the **maxprobe full-drafter PDI that established the M7 silicon baseline** (the image on the
- report:MONO_BUILD_FIX_RECIPE_20260520.md | 2026-05-20 | historical | Mono Build Fix Recipe (2026-05-20)
  <repo>/reports/MONO_BUILD_FIX_RECIPE_20260520.md
  Companion to `reports/CRITICAL_PDI_MISMATCH_20260520.md`. The previous /build_all.sh run was building V3 split, not the spatial-fix `dummy_layer`. Here is the exact path to build the spatial-fix kerne
- report:NEXT_STEPS_LADDER.md | 2026-09-02 | current | EAGLE4 V80 Drafter — Next-Steps + Integration Ladder [0xE4B71101]
  <repo>/reports/NEXT_STEPS_LADDER.md
  **Date:** 2026-06-14 · **Author:** orchestrator synthesis (read-only; no builds/edits)
- report:ONBOARDING_100_DOCS.md | 2026-07-23 | historical | ONBOARDING — the 100 markdowns a brand-new instance should read [0xE4B71111,0xE4B71115,0xE4B71118,0xE4B7111B]
  <repo>/reports/ONBOARDING_100_DOCS.md
  *These encode failures that each cost a day or more. Pattern-match against them first.*
- report:ORCHESTRATOR_BLACKOUT_LOG.md | 2026-10-22 | current | Orchestrator Blackout Log [0xE4B71159,0xE4B71199,0xE4B711A9,0xE4B711AE...]
  <repo>/reports/ORCHESTRATOR_BLACKOUT_LOG.md
  Append-only. Written by `scripts/orchestrator_blackout_watchdog.sh`, which runs OUTSIDE any
- report:OSPREY_CHECKLIST.md | 2026-08-21 | historical | OSPREY — the EagleFIT/bitnet_721 universal image campaign (build-host-M) [0xE4B71118,0xE4B71128]
  <repo>/reports/OSPREY_CHECKLIST.md
  **Named, isolated campaign** (user directive 2026-07-29). Do NOT confound with the GPTQ lineage
- report:PARAMETER_REALITY_AUDIT.md | 2026-07-31 | historical | PARAMETER REALITY AUDIT — what the RTL actually does vs what we tell ourselves [0xE4B7110F,0xE4B71115,0xE4B71117]
  <repo>/reports/PARAMETER_REALITY_AUDIT.md
  **Date:** 2026-07-21 · **Scope:** READ-ONLY audit of three forks. No edits, no card access, no build interaction.
- report:PDI11_DEPTHN_URAM_PLAN.md | 2026-06-11 | historical | PDI#11 — Depth-N + URAM-Widen: Executable Checklist
  <repo>/reports/PDI11_DEPTHN_URAM_PLAN.md
  **Goal:** ship BOTH accuracy-preserving levers as fast as possible — (A) multi-depth tree
- report:PDI6_BUILD_EXECUTION_RUNBOOK_20260527.md | 2026-05-27 | historical | PDI#6 build-execution runbook: 6 steps synth-gate→survival→re-ooc→build→archive→flash, each gated, a
  <repo>/reports/PDI6_BUILD_EXECUTION_RUNBOOK_20260527.md
  **Date:** 2026-05-27
- report:PDI6_CHECKKLIST.md | 2026-06-05 | historical | PDI6_CHECKKLIST.md
  <repo>/reports/PDI6_CHECKKLIST.md
  
- report:PDI6_CHECKLIST.md | 2026-06-05 | historical | PDI#6 — Live Goal Checklist (driven by the 5-min cron)
  <repo>/reports/PDI6_CHECKLIST.md
  **Updated each cron tick.** Goal: a flash-ready, xsim-golden-bit-exact, URAM-fitting, positive-WNS @125 MHz
- report:PDI6_FEEDBACK_DECISION.md | 2026-06-05 | historical | PDI#6 FEEDBACK — Live Decision Doc
  <repo>/reports/PDI6_FEEDBACK_DECISION.md
  **Updated:** 2026-05-30 (post-WO-035 RCA). **Goal:** a flashable GOLDEN (with-G8) PDI#6 — post-route Common WNS ≥ 0, bake clean, surfaced for user cold-AC.
- report:PDI71_ARTIFACT_AUDIT.md | 2026-06-06 | historical | PDI#7.1 Pre-Flash Artifact Audit (Fork-B)
  <repo>/reports/PDI71_ARTIFACT_AUDIT.md
  Build: `AVED-..._b/build` (sourced `create_bd_design_layout_b.tcl`); kernel IP `top_eagle4_drafter_lb_0_0`. Synth host lab-host, 2026-06-05 19:47. Route complete (routed.dcp 06-06 01:14, WNS 0.000); bi
- report:PDI71_DMA_AUDIT.md | 2026-06-06 | historical | PDI71 DMA-Layer Line-by-Line Audit (2026-06-06)
  <repo>/reports/PDI71_DMA_AUDIT.md
  Scope: `EAGLE/v80_test_suite/qdma_utils.py`, `03_setup_qdma.sh`, and the DMA paths in
- report:PDI71_HOST_AUDIT.md | 2026-06-06 | historical | PDI#7.1 Host-Driver Line-by-Line Audit (2026-06-06)
  <repo>/reports/PDI71_HOST_AUDIT.md
  Scope: the paths that run right after tonight's PDI#7.1 flash —
- report:PDI71_REGMAP_FINAL.md | 2026-06-06 | historical | PDI#7.1 Register-Map Final Cross-Check
  <repo>/reports/PDI71_REGMAP_FINAL.md
  **Date:** 2026-06-06 · **Verdict: GO (with one host guard — see §Verdict)**
- report:PDI72_BUILD_TICKS.md | 2026-06-07 | historical | PDI72_BUILD_TICKS.md
  <repo>/reports/PDI72_BUILD_TICKS.md
  04:40 | build=packaging-synth | errors=0 | m7=pending | launcher in pre-package util synth (nrm_buf RAM-dissolve warnings benign/pre-existing)
- report:PDI7_BUILD_LAUNCH.md | 2026-06-05 | historical | PDI#7 Build Launch — 2026-06-04
  <repo>/reports/PDI7_BUILD_LAUNCH.md
  **Status:** LAUNCHED, launch gate PASS. Orchestrator monitors to post-route.
- report:PDI7_BUILD_READY.md | 2026-06-05 | historical | PDI#7 BUILD-READY — G8 watchdog hardening, 100 MHz golden (2026-06-04)
  <repo>/reports/PDI7_BUILD_READY.md
  1. **Architecture** — Full G8-spliced EAGLE4 drafter top (`eagle4_drafter`, fork
- report:PDI8_DBG_BUILD.md | 2026-06-05 | historical | PDI#8 FULL-PROBE (v1+v2+v3 debug) — queued build
  <repo>/reports/PDI8_DBG_BUILD.md
  Date 2026-06-05 · Fork B · source = `verilog/forks/pdi6_s7c_g8` HEAD (G8 golden + dbg v1/v2/v3).
- report:PDI8_PREVIEW_BUILD.md | 2026-06-05 | historical | PDI#8-preview — "latest-everything" build (fork A), parallel to PDI#7 (fork B)
  <repo>/reports/PDI8_PREVIEW_BUILD.md
  **Date:** 2026-06-05
- report:PDI9_POSTFLASH_RUNBOOK.md | 2026-06-07 | historical | PDI#9 Post-Flash Runbook (prepared 2026-06-06 21:0x, pre-route)
  <repo>/reports/PDI9_POSTFLASH_RUNBOOK.md
  The image being built is the FIRST from sim-golden RTL (cheap-set 8/8 ordered, ROUND-15).
- report:PDI9_PREVIEW_DBGV2.md | 2026-06-05 | historical | PDI#9-preview ← debug-block-v2 merge (COORDINATOR)
  <repo>/reports/PDI9_PREVIEW_DBGV2.md
  **Goal:** ship the optimal PDI#9-preview build WITH full intermediate-layer debug
- report:PDI_0xE4B7110F_BATCH_TESTPLAN.md | 2026-07-10 | historical | PDI 0xE4B7110F (TREE-BATCH B=2) — what it is, test plan, and the analysis backlog [0xE4B7110E,0xE4B7110F]
  <repo>/reports/PDI_0xE4B7110F_BATCH_TESTPLAN.md
  **What it DOES:** adds `BODY_BATCH_B=2` — the body GEMV array reads each 2-bit weight word from
- report:PDI_BUILD_INDEX.md | 2026-07-10 | historical | PDI BUILD INDEX — latest generation (compaction-survival index) [0xE4B71107,0xE4B71108,0xE4B71109,0xE4B7110A...]
  <repo>/reports/PDI_BUILD_INDEX.md
  **User-requested 2026-07-06** ("index the latest three pdi builds with reason, code change, file
- report:PDI_DRIVE_PLAN_20260827.md | 2026-08-27 | historical | 📌 PDI DRIVE PLAN — the build→assess→optimize loop to a flashable FS=1 image (PINNED 2026-08-27) [0xE4B71128,0xE4B71167]
  <repo>/reports/PDI_DRIVE_PLAN_20260827.md
  **Standing objective (user-ordered): drive to a SUCCESSFUL PDI — positive WNS setup AND hold,
- report:PROFESSOR_STATUS_20260628.md | 2026-06-28 | historical | EAGLE4 2-bit Drafter on Alveo V80 — Status for Review (2026-06-28)
  <repo>/reports/PROFESSOR_STATUS_20260628.md
  The hand-written SystemVerilog 2-bit speculative-decoding drafter now runs **bit-exact through the full transformer body on real V80 silicon** and **decodes end-to-end inside SGLang on hardware**. The
- report:PROJECT_ASSESSMENT_20260609.md | 2026-06-09 | historical | EAGLE4 V80 2-bit Drafter — Independent Project Assessment
  <repo>/reports/PROJECT_ASSESSMENT_20260609.md
  **Reviewer:** Fresh outside-eyes pass (Grand Master PL, no prior involvement in this fork).
- report:PROVEN_SPEED_LEDGER.md | 2026-07-31 | historical | PROVEN_SPEED_LEDGER.md — EAGLE4 V80 Drafter Authoritative Per-Stage Cycle×Clock Ledger
  <repo>/reports/PROVEN_SPEED_LEDGER.md
  **Date:** 2026-06-13 (rebuilt from disk; the silicon↔xsim reconciliation is now CLOSED)
- report:QSFP_DRAFTER_COLLABORATION_HANDOFF.md | 2026-09-14 | current | QSFP transport → full-drafter collaboration handoff [0xE4B71179]
  <repo>/reports/QSFP_DRAFTER_COLLABORATION_HANDOFF.md
  **September 14 recovery entry:** [Current multi-file context dossier](<repo>/reports/qsfp_context_recovery_20260914/00_START_HERE.md) records 
- report:QSFP_LINK_CAMPAIGN_INDEX.md | 2026-09-11 | current | QSFP INTER-FPGA LINK — MASTER INDEX
  <repo>/reports/QSFP_LINK_CAMPAIGN_INDEX.md
  **One file linking every repo, experiment, observation, script and source artifact of the QSFP link campaign.**
- report:REDISCOVERED_STATE_20260614.md | 2026-06-14 | historical | REDISCOVERED STATE — 2026-06-14 (post-compaction context recovery) [0xE4B71101]
  <repo>/reports/REDISCOVERED_STATE_20260614.md
  This file is the authoritative "where we actually are" snapshot, rebuilt by reading the
- report:RESULTS_AND_REPRODUCE_20260723.md | 2026-07-23 | historical | 2026-07-23 — RESULTS + FULL REPRODUCE RUNBOOK + depth-N end-to-end plan [0xE4B71118,0xE4B71119,0xE4B7111B]
  <repo>/reports/RESULTS_AND_REPRODUCE_20260723.md
  First image that verifiably contains `HEAD_PAR=6` (`check_strap_netlist.sh` PASS; netlist has
- report:SCALE_CONSTFOLD_FEASIBILITY.md | 2026-06-05 | historical | Scale Const-Fold Feasibility — bake per-group GPTQ scales into LUT-ROM (PDI#6, V80)
  <repo>/reports/SCALE_CONSTFOLD_FEASIBILITY.md
  **Date:** 2026-06-03   **Part:** xcv80-lsva4737-2MHP-e-S   **Tool:** Vivado 2025.1 OOC synth
- report:SESSION_20260722_MASTER_LOG.md | 2026-07-22 | historical | 2026-07-22 — MASTER LOG: progress, evidence, and reproducible steps [0xE4B71115,0xE4B71118,0xE4B71119,0xE4B7111B]
  <repo>/reports/SESSION_20260722_MASTER_LOG.md
  **Measured per-forward (N=60, ±0.1%), `reports/results/monolith_v2_campaign_data.json`:**
- report:SESSION_CONSOLIDATED_20260823.md | 2026-08-23 | historical | EAGLE4 — CONSOLIDATED SESSION STATE, 2026-08-22 → 2026-08-23 03:3x EDT [0xE4B71128,0xE4B71143,0xE4B71155,0xE4B71157...]
  <repo>/reports/SESSION_CONSOLIDATED_20260823.md
  Anti-context-loss digest. Every number here is MEASURED unless tagged DERIVED / NOT-MEASURED.
- report:SESSION_LEARNINGS_20260614.md | 2026-06-14 | historical | SESSION LEARNINGS — 2026-06-14 (anti-compaction: read before re-deciding anything) [0xE4B71101]
  <repo>/reports/SESSION_LEARNINGS_20260614.md
  Dense, evidence-linked record so the NEXT context does NOT repeat this session's wrong impressions.
- report:SESSION_LEDGER_20260709_10.md | 2026-07-09 | historical | SESSION EXPERIMENTAL LEDGER — 2026-07-09/10 (the WIDE→batch→wave arc) [0xE4B7110E,0xE4B7110F]
  <repo>/reports/SESSION_LEDGER_20260709_10.md
  From GREEN-gate RDONE stamps (xsim clock 2 ns verified: tree window == documented 312,631 cyc);
- report:SESSION_LEDGER_20260711_13.md | 2026-07-11 | historical | SESSION LEDGER 2026-07-11 → 07-13 — jewel arc + reclock saga + speed wave + 3-PDI drive (ANTI-COMPAC [0xE4B7110F,0xE4B71110,0xE4B71111,0xE4B71112]
  <repo>/reports/SESSION_LEDGER_20260711_13.md
  "Prepare the PDI image WITHOUT me asking … don't waste 8 hours waiting on a meaningless confirmation … get me the PDI image to actually consolidate any progress." ⇒ every gated-RTL wave converges into
- report:SGLANG_3B_2BIT_FROM_CHAT_HISTORY.md | 2026-06-05 | historical | SGLang Llama-3.2-3B + EAGLE4 2-bit — Forensic Reconstruction From Chat History
  <repo>/reports/SGLANG_3B_2BIT_FROM_CHAT_HISTORY.md
  **Date written:** 2026-04-26
- report:SGLANG_LLAMA3B_2BIT_RUNBOOK.md | 2026-06-05 | historical | SGLang Runbook — Llama-3.2-3B + EAGLE4 2-bit GPTQ Drafter
  <repo>/reports/SGLANG_LLAMA3B_2BIT_RUNBOOK.md
  **Date:** 2026-04-25
- report:SGLANG_MODE1_HLS_HOOK_CONFIRMED_20260427.md | 2026-04-27 | historical | SGLang `--speculative-exec-mode 1` HLS Hook Empirical Confirmation
  <repo>/reports/SGLANG_MODE1_HLS_HOOK_CONFIRMED_20260427.md
  **Date:** 2026-04-27 09:30 EDT
- report:SGLANG_TREE_FLOW_REFERENCE.md | 2026-07-22 | historical | SGLang Tree Flow Reference — the drafter algorithm the FPGA tree FSM implements [0xE4B71115]
  <repo>/reports/SGLANG_TREE_FLOW_REFERENCE.md
  **Created** 2026-07-22 · **Companion to** `reports/SESSION_20260722_MASTER_LOG.md`
- report:SGLANG_TREE_KV_SURFACE.md | 2026-07-21 | historical | SGLang Tree + KV Surface — the complete inventory of what the FPGA would have to take over [0xE4B71115]
  <repo>/reports/SGLANG_TREE_KV_SURFACE.md
  **Date:** 2026-07-21
- report:SILICON_PROOF_WORKLOAD_PDI6.md | 2026-06-05 | historical | Silicon-Proof Workload — PDI#6 (real-drafter bit-exact on V80)
  <repo>/reports/SILICON_PROOF_WORKLOAD_PDI6.md
  **Created 2026-05-25.** Owner: orchestrator (agents WRITE, orchestrator REVIEWS). Goal: a REAL
- report:SOTA_URAM_BW_WALL.md | 2026-09-02 | current | SOTA SURVEY — Beating the URAM Weight-Read Bandwidth Wall
  <repo>/reports/SOTA_URAM_BW_WALL.md
  **Date:** 2026-06-14
- report:SPEEDWAVE_0xE4B71111_SILICON_REPORT.md | 2026-07-14 | historical | SPEEDWAVE (PDI 0xE4B71111) — Build, Flash & Silicon-Validation Report [0xE4B71102,0xE4B71110,0xE4B71111]
  <repo>/reports/SPEEDWAVE_0xE4B71111_SILICON_REPORT.md
  SPEEDWAVE = the silicon-validated **reclock66** baseline (0xE4B71110) + three speed levers, each
- report:SPEED_ATTN_ANALYSIS.md | 2026-06-13 | historical | SPEED_ATTN_ANALYSIS — best accuracy-neutral G5 attention speedup for the grand build
  <repo>/reports/SPEED_ATTN_ANALYSIS.md
  **Date:** 2026-06-13
- report:SPEED_BUILD_WORKORDERS.md | 2026-09-02 | current | SPEED-BUILD WORK ORDERS — live ledger (resume-after-compaction) [0xE4B7111B,0xE4B7111D]
  <repo>/reports/SPEED_BUILD_WORKORDERS.md
  **GOAL (terminal):** launch ONE consolidated **best-speed PDI** *in addition to* build #1 (correctness).
- report:SPEED_RESOURCE_ENVELOPE.md | 2026-06-13 | historical | SPEED_RESOURCE_ENVELOPE.md — How much MORE parallelism the grand-speed build can add within the V80 
  <repo>/reports/SPEED_RESOURCE_ENVELOPE.md
  **Date:** 2026-06-13
- report:SPEED_RESPIN_0x1126_WORKORDER.md | 2026-07-29 | historical | SPEED RESPIN 0xE4B71126 — CONSOLIDATED WORKORDER (drafted 2026-07-29) [0xE4B71118,0xE4B71125,0xE4B71126]
  <repo>/reports/SPEED_RESPIN_0x1126_WORKORDER.md
  **Sequencing (hard gate): fires ONLY after the 0xE4B71125 flash verdict** (accept campaign on silicon;
- report:SPEED_TREE_OVERLAP_ANALYSIS.md | 2026-06-13 | historical | SPEED ANALYSIS — Tree Further-Cut + Inter-Stage Overlap (2026-06-13)
  <repo>/reports/SPEED_TREE_OVERLAP_ANALYSIS.md
  **Scope:** ANALYSIS ONLY (no build/xsim/OOC run). Two cycle levers owned here:
- report:SPEED_URAM_BW_COMPUTE.md | 2026-09-02 | current | SPEED — URAM READ BANDWIDTH & RAW COMPUTE: measured + ranked improvements
  <repo>/reports/SPEED_URAM_BW_COMPUTE.md
  **Date:** 2026-06-13
- report:STATE.md | 2026-06-14 | historical | STATE.md
  <repo>/reports/STATE.md
  - [08:39] write_device_image DONE (peak 54G released, free=32G). I VERIFIED timing myself: post_route_phys_opt_report_timing_summary_0.rpt = "All user specified timing constraints are met" (setup+hold
- report:STATUS_AND_EXPERIMENTS_20260823.md | 2026-08-23 | historical | EAGLE4 — WHERE WE STAND, AND EVERY EXPERIMENT RUN (2026-08-23) [0xE4B71127,0xE4B71158]
  <repo>/reports/STATUS_AND_EXPERIMENTS_20260823.md
  **Written by the orchestrator on user request. Every figure is labelled [MEASURED] / [DERIVED] /
- report:STORAGE_RELOCATION_INDEX.md | 2026-06-05 | historical | Storage Relocation Index — 2026-05-20
  <repo>/reports/STORAGE_RELOCATION_INDEX.md
  Disk hit 98% full, killing 2 background builds (mono PDI 158126 + G2 HLS 173981).
- report:TODO_NEXT_20260902.md | 2026-09-02 | current | TODO — ranked next actions, 2026-09-02 [0xE4B71110,0xE4B71143,0xE4B71159,0xE4B7115B...]
  <repo>/reports/TODO_NEXT_20260902.md
  Written at the close of the "why is our compute slow" investigation. Every item carries WHY, HOW,
- report:TOKEN_LATENCY_MODEL.md | 2026-09-02 | current | TOKEN_LATENCY_MODEL.md — EAGLE4 V80 Drafter: End-to-End tok/s, Not Per-Forward [0xE4B71143]
  <repo>/reports/TOKEN_LATENCY_MODEL.md
  **Date:** 2026-06-13
- report:TONIGHT_EVIDENCE_LEDGER.md | 2026-06-14 | historical | Tonight's Operation — Master Evidence Ledger (2026-06-14)
  <repo>/reports/TONIGHT_EVIDENCE_LEDGER.md
  Synthesis of 5 read-only consolidation passes. Every claim tagged by tier:
- report:TONIGHT_SUMMARY_20260721.md | 2026-07-21 | historical | EAGLE4 FPGA drafter — status as of 2026-07-21 [0xE4B71111,0xE4B71115]
  <repo>/reports/TONIGHT_SUMMARY_20260721.md
  Both bases move ~35% — the gain is not an artifact of benchmark choice.
- report:TREE_CAMPAIGN_MASTER_INDEX.md | 2026-07-25 | historical | Tree-FSM Campaign — Master Index (2026-07-24 → 07-25) [0xE4B7111E]
  <repo>/reports/TREE_CAMPAIGN_MASTER_INDEX.md
  **Goal:** depth-N tree drafter, verified end-to-end against SGLang, built to a flashable PDI.
- report:TREE_FSM_DESIGN.md | 2026-07-24 | historical | TREE FSM — Design + Staged Implementation Plan (DESIGN DOCUMENT, no RTL yet) [0xE4B71117]
  <repo>/reports/TREE_FSM_DESIGN.md
  **Created** 2026-07-21 · **Status** DESIGN FOR USER REVIEW — no production RTL written
- report:TREE_MECHANISM_AND_LIVE_VERIFICATION.md | 2026-07-24 | historical | The Tree Mechanism & Tonight's Live Verification
  <repo>/reports/TREE_MECHANISM_AND_LIVE_VERIFICATION.md
  **2026-07-24 ~20:00 · written for review while sitting together · corrections welcome — this doc is my current understanding; if any ❓ item is wrong, that changes RTL/goldens**
- report:TREE_VALIDATION_STATUS.md | 2026-07-24 | historical | TREE / DEPTH-N VALIDATION — status, evidence, and remaining gaps [0xE4B71118,0xE4B71119,0xE4B7111B]
  <repo>/reports/TREE_VALIDATION_STATUS.md
  The depth-N selection FSM is proven against real GPU-decode surfaces through two levels, and the
- report:UNITPDI_LADDER_PREP.md | 2026-06-14 | historical | UNIT-PDI Ladder Prep — U-NCH (N=4 striped gather) + U-WIDE (512b m_axi_g8)
  <repo>/reports/UNITPDI_LADDER_PREP.md
  **Date:** 2026-06-14
- report:UNTRIED_HLS_PRAGMAS_INVENTORY.md | 2026-06-05 | historical | Untried HLS Pragmas Inventory — EAGLE4 V80 (Vitis HLS 2025.1)
  <repo>/reports/UNTRIED_HLS_PRAGMAS_INVENTORY.md
  **Scope:** pragmas / directives / config knobs we have NOT yet exercised in silicon-validated kernels (`eagle4_drafter_dummy_vocab_lmdown_uramqw.cpp`, `eagle4_drafter_dummy_layer.cpp`, `eagle4_lm_head
- report:URAM_COOKBOOK_VERSAL_V80.md | 2026-06-05 | historical | URAM Cookbook — AMD Versal V80 (xcv80-lsva4737-2MHP-e-S), Vitis HLS 2025.1
  <repo>/reports/URAM_COOKBOOK_VERSAL_V80.md
  Scope: unexercised knobs. Citations: UG953, UG1399 (Vitis HLS 2025.1), UG573 (URAM288 primitive), our two sources.
- report:URAM_EXPERIMENTS_EAGLE4.md | 2026-06-05 | historical | URAM Experiments — EAGLE4 FPGA Drafter
  <repo>/reports/URAM_EXPERIMENTS_EAGLE4.md
  Date: 2026-05-20. Author: fpga-hls-perf-architect.
- report:URA_INTERVIEW_TYLER_CHEN.md | 2026-07-31 | historical | URA Interview Kit — Tyler Chen
  <repo>/reports/URA_INTERVIEW_TYLER_CHEN.md
  **Role:** Undergraduate Research Assistant (URA) — "Speculative Decoding and Inference Acceleration for LLMs"
- report:V3_BANKED_CHAIN_REFACTOR_PLAN.md | 2026-05-07 | historical | v3_banked refactor plan for chain GEMVs
  <repo>/reports/V3_BANKED_CHAIN_REFACTOR_PLAN.md
  **Status:** Design plan — not yet implemented. Next loop session can pick this up.
- report:VERILOG_STARTER_KIT.md | 2026-06-05 | historical | Verilog Experiment Pathway — Starter Kit
  <repo>/reports/VERILOG_STARTER_KIT.md
  **Audience:** an agent (or human) starting Verilog work on this V80 project for the first time. Assumes you know Verilog/SystemVerilog syntax + Vivado basics. Does NOT assume you know our AVED shell, 
- report:VERILOG_SYSTOLIC_PROPOSAL_FRESH_AGENT.md | 2026-06-05 | historical | Verilog-Systolic Redesign — Feasibility Proposal
  <repo>/reports/VERILOG_SYSTOLIC_PROPOSAL_FRESH_AGENT.md
  **Author:** fresh architecture-exploration agent (no prior project context)
- report:W_DISPATCHER_0x1129_WORKORDER.md | 2026-07-31 | historical | W-DISPATCHER WORKORDER — 0xE4B71129 headline (drafted 2026-07-30, coordinator ruling) [0xE4B71118,0xE4B71127,0xE4B71129]
  <repo>/reports/W_DISPATCHER_0x1129_WORKORDER.md
  **Why this exists:** the 0x1126 sprint's specialist audit found the W-sibling DISPATCHER does not exist

## repo_doc (59)

- doc:ARCH_DIRECTIONS.md | 2026-08-20 | historical | EAGLE4 on V80 — Architecture Directions [0xE4B71128]
  <repo>/docs/ARCH_DIRECTIONS.md
  Most arguments in this project have conflated them. Separating them is the main contribution of this note.
- doc:ASIC_PER_PHASE_ARCHITECTURE.md | 2026-07-28 | historical | EAGLE4 Drafter — Per-Phase Spatial ASIC Architecture [0xE4B7110F,0xE4B71111,0xE4B71118,0xE4B71120]
  <repo>/docs/ASIC_PER_PHASE_ARCHITECTURE.md
  Dimensions, all verified against RTL source (not against documentation):
- doc:AUTO_FEEDBACK_LOOP.md | 2026-08-25 | historical | THE AUTOMATIC FEEDBACK LOOP — golden -> mismatch -> fix -> PDI -> proof [0xE4B71118]
  <repo>/docs/AUTO_FEEDBACK_LOOP.md
  2026-08-24. **This is an INDEX, not a replacement.** Every stage below already has
- doc:AXI_STREAM_INTEGRATION_PLAN.md | 2026-07-06 | historical | AXI-Stream Integration Plan (V6_axistream)
  <repo>/docs/AXI_STREAM_INTEGRATION_PLAN.md
  **Date:** 2026-05-18
- doc:BUILD_LAUNCH_README.md | 2026-08-02 | historical | How to Launch a PDI Build (lab-host / build-host-M) — Onboarding README [0xE4B71118,0xE4B71124,0xE4B71127]
  <repo>/docs/BUILD_LAUNCH_README.md
  **Audience:** a new team member with SSH access to **build-host-M (<ip>, build-only box)**
- doc:BUILD_VERIFICATION_CHAIN_PLAN.md | 2026-08-22 | historical | EAGLE4 V80 — Sibling, Parallel, and Next-PDI Builds: the Verification Chain [0xE4B71118,0xE4B71155]
  <repo>/docs/BUILD_VERIFICATION_CHAIN_PLAN.md
  **Date 2026-08-21 · Status: LIVING PLAN.** This is the operating document for every PDI from here on. It
- doc:CTO_DEBUG_MINDSET.md | 2026-07-06 | historical | CTO DEBUG MINDSET — how the silicon tree was made exact, and how to think when I'm not here
  <repo>/docs/CTO_DEBUG_MINDSET.md
  **2026-07-06.** This documents the reasoning chain that cracked the G8RESC bug — five silicon
- doc:DEEP_TREE_SHIFT_PLAN_20260822.md | 2026-08-22 | historical | ENGINEERING SHIFT: design the FPGA to recreate the depth-15 / HumanEval regime
  <repo>/docs/DEEP_TREE_SHIFT_PLAN_20260822.md
  2026-08-22 23:xx — plan of record for the overnight run. Author: orchestrator.
- doc:DESIGN_G8_HBM32.md | 2026-07-06 | historical | DESIGN_G8_HBM32 — all-32-pseudo-channel lm-head gather
  <repo>/docs/DESIGN_G8_HBM32.md
  - **It was never DRAM latency.** Flashed scalar G8 = 6,144 ARs × ~514 cyc/AR, and the 514
- doc:DESIGN_KV_ONCHIP_DEPTHN.md | 2026-07-11 | historical | DESIGN — Fully on-chip KV cache for depth-N trees, with smart FF intermediates
  <repo>/docs/DESIGN_KV_ONCHIP_DEPTHN.md
  Depth-N needs each tree branch to attend over **its own ancestor context**:
- doc:DESIGN_PERSTAGE_PIPELINE.md | 2026-07-05 | historical | Design: Per-Stage GEMV Arrays → Full Node-Pipelined Dataflow Drafter [0xE4B71107]
  <repo>/docs/DESIGN_PERSTAGE_PIPELINE.md
  **Status:** DESIGN (not scheduled — gated behind G8RESC fix + WIDE gather, see §10)
- doc:DESIGN_PS_JEWEL.md | 2026-07-06 | historical | DESIGN_PS_JEWEL — the per-stage pipelined drafter (final parallel architecture)
  <repo>/docs/DESIGN_PS_JEWEL.md
  **Intent (user, 2026-07-06):** the most ideal parallel hardware — individual per-stage GEMM
- doc:DMA_DRIVER_GUIDE.md | 2026-07-06 | historical | DMA Driver Guide — Host ↔ V80 ↔ H200 GPU (EAGLE4 / QDMA)
  <repo>/docs/DMA_DRIVER_GUIDE.md
  How data and control move between the host, the AMD Alveo V80 (Versal HBM,
- doc:EAGLE4_COMPUTE_AND_KV_TRACKING.md | 2026-07-01 | historical | EAGLE4 Drafter: Compute Pipeline and KV-Cache Tracking for One-Shot Silicon
  <repo>/docs/EAGLE4_COMPUTE_AND_KV_TRACKING.md
  **Scope.** This document specifies, at an implementation-verifiable level, (1) the full
- doc:EAGLE4_MAINLINE_RESTART_GUIDE.md | 2026-07-06 | historical | EAGLE4 FPGA Mainline Restart Guide
  <repo>/docs/EAGLE4_MAINLINE_RESTART_GUIDE.md
  **Date written:** 2026-05-13
- doc:EXPERIMENTS_URAM_BANDWIDTH.md | 2026-07-07 | historical | EXPERIMENTS_URAM_BANDWIDTH — closing the 7.4%→100% URAM port-utilization gap
  <repo>/docs/EXPERIMENTS_URAM_BANDWIDTH.md
  **Idea:** the SPLIT2 cascade fix (07-06) gave every big bank TWO independent URAM cascades
- doc:FABLE_MASTER_PLAN.md | 2026-09-02 | current | FABLE_MASTER_PLAN — the architecture + experiment plan of record [0xE4B71109]
  <repo>/docs/FABLE_MASTER_PLAN.md
  This is the plan as it stood when authored — a reference, not a live tracker. Where a claim was
- doc:FPGA_SILICON_BRINGUP_BEST_PRACTICES.md | 2026-07-01 | historical | Building Correct Silicon from Verilog: A Field-Earned Methodology
  <repo>/docs/FPGA_SILICON_BRINGUP_BEST_PRACTICES.md
  **Thesis.** On a large multi-SLR FPGA, a simulation-green, timing-met bitstream is *not*
- doc:FUNDAMENTAL_COMPUTE_OPTIMAL_DESIGN.md | 2026-07-06 | historical | FUNDAMENTAL COMPUTE PRINCIPLES → OPTIMAL DESIGN — EAGLE4 drafter on V80 [0xE4B7110C]
  <repo>/docs/FUNDAMENTAL_COMPUTE_OPTIMAL_DESIGN.md
  1. **Batch-1 Llama decode = a GEMV chain.** Every 2-bit weight is read exactly ONCE per forward
- doc:GPU_ACCEPT_GOLDEN_COMMAND.md | 2026-08-25 | historical | 📌 PINNED — THE GPU ACCEPT REFERENCE COMMAND (user-supplied 2026-08-24 06:01)
  <repo>/docs/GPU_ACCEPT_GOLDEN_COMMAND.md
  **This is the command. Do not reconstruct it, do not substitute paths, do not "improve" it.**
- doc:GPU_ACCEPT_RUNBOOK.md | 2026-08-27 | historical | GPU ACCEPT RUNBOOK — how to reproduce every accept number, from a cold shell [0xE4B7115B]
  <repo>/docs/GPU_ACCEPT_RUNBOOK.md
  **Why this exists:** every accept figure this project quotes was produced by one of two harnesses that
- doc:GPU_BASELINE_MEASUREMENT_PLAN.md | 2026-07-06 | historical | GPU Baseline Measurement Plan for EAGLE4 FPGA Paper
  <repo>/docs/GPU_BASELINE_MEASUREMENT_PLAN.md
  **Date:** 2026-05-17
- doc:HANDOFF_PHASE3_INTEGRATION.md | 2026-06-05 | historical | EAGLE4 Verilog Drafter — Phase 3 Handoff (PDI #2 → PDI #3)
  <repo>/docs/HANDOFF_PHASE3_INTEGRATION.md
  **Last edited:** 2026-05-22
- doc:HARNESS.md | 2026-08-24 | historical | The harness, consolidated
  <repo>/docs/HARNESS.md
  2026-08-24. One page for "what checks exist, which ones actually run, and what
- doc:INNOVATION_BOARD_20260924.md | 2026-09-24 | current | Innovation board — pair speed (opened 2026-09-24 03:40, orchestrator HQ) [0xE4B711CE,0xE4B711CF,0xE4B711D2]
  <repo>/docs/INNOVATION_BOARD_20260924.md
  User directive 03:30: "focus is on overhauling speed improvements on various creative directions and factors of the
- doc:INTEGRATION_HANG_ROOT_CAUSE_ANALYSIS.md | 2026-07-06 | historical | Integration-Hang Root Cause Analysis
  <repo>/docs/INTEGRATION_HANG_ROOT_CAUSE_ANALYSIS.md
  **Date:** 2026-05-17
- doc:JEWEL_SILICON_PDI_LADDER.md | 2026-09-02 | current | JEWEL SILICON PDI LADDER — direction-proving PDIs toward the per-stage pipelined drafter [0xE4B7110E]
  <repo>/docs/JEWEL_SILICON_PDI_LADDER.md
  **Principle (user 2026-07-09):** don't build one giant jewel PDI at the end. Build a LADDER of small
- doc:KV_CACHE_PREFIX_BUG_REPORT.md | 2026-06-05 | historical | KV Cache Prefix Bug — Investigation Report and Patch Design
  <repo>/docs/KV_CACHE_PREFIX_BUG_REPORT.md
  Author: FPGA HLS tech-lead agent
- doc:MARKDOWN_TAXONOMY.md | 2026-07-06 | historical | MARKDOWN TAXONOMY — long-term organization of the ~1,691 repo markdowns
  <repo>/docs/MARKDOWN_TAXONOMY.md
  **Written 2026-07-06 (Phase 1 of the phased+safety-gated reorg).** Executed same day (Phase 2) for
- doc:MESH_DISPATCH.md | 2026-09-24 | current | Mesh dispatch — memory-reservation admission for lab-host / build-host-A / laptop (2026-09-23)
  <repo>/docs/MESH_DISPATCH.md
  **Why:** three overruns on 2026-09-23 (laptop OOM-killed a route and later thrashed to a reboot; lab-host hit 112/125 GB + swap).
- doc:MODULE_ATLAS.md | 2026-06-11 | historical | EAGLE4 Drafter — MODULE ATLAS
  <repo>/docs/MODULE_ATLAS.md
  Miniature computational illustrations (operations + dataflow) for every module on the
- doc:MUST_PASS_VERIFICATION_LIST.md | 2026-10-22 | current | MUST-PASS VERIFICATION LIST: test the built artifact, not the RTL (user-mandated 2026-10-22) [0xE4B71118]
  <repo>/docs/MUST_PASS_VERIFICATION_LIST.md
  **Why this exists.** On 2026-10-22, the first Qwen silicon test of 0xE4B7129B failed after about three days of building and routing. The dual-model RTL passed xsim, lint and the Llama back-compat test
- doc:NEXT_IMAGE_DESIGN_RULES.md | 2026-09-22 | current | 🔴 DESIGN RULES FOR THE NEXT PDI IMAGE — **NOTHING MAY BE UN-CLEARABLE**
  <repo>/docs/NEXT_IMAGE_DESIGN_RULES.md
  **USER DIRECTIVE 2026-09-22: "for next pdi image, do not design stuff that can't be cleared."**
- doc:OPEN_QUESTIONS_FUTURE_WORK.md | 2026-07-06 | historical | Open Questions and Future Work — Consolidated Index
  <repo>/docs/OPEN_QUESTIONS_FUTURE_WORK.md
  **Compiled:** 2026-05-18
- doc:OPERATING_PROTOCOL.md | 2026-07-06 | historical | EAGLE4 FPGA — Operating Protocol
  <repo>/docs/OPERATING_PROTOCOL.md
  **The "how we operate" bible.** This codifies the working DISCIPLINES — the habits that
- doc:PACKAGING_SCRIPT_CHAIN.md | 2026-08-19 | historical | PACKAGING SCRIPT CHAIN — RTL → packaged EAGLE4 IP [0xE4B71118,0xE4B71127,0xE4B71129]
  <repo>/docs/PACKAGING_SCRIPT_CHAIN.md
  **Audience:** Tyler (new team member) who wants to *read the source* of every stage.
- doc:PANDOC_INSTALL.md | 2026-05-18 | historical | Pandoc + LaTeX Install Guide (EAGLE4 V80 Paper)
  <repo>/docs/PANDOC_INSTALL.md
  This document is the install pre-flight for `scripts/build_paper.sh`, which
- doc:PAPER_COMPILED.md | 2026-07-06 | historical | Title: A Bit-Exact 2-bit EAGLE4 Speculative-Decoding Drafter for AMD Versal V80 — Honest Per-Forward
  <repo>/docs/PAPER_COMPILED.md
  **Title option A (recommended):** *A Bit-Exact 2-bit EAGLE4 Drafter on AMD Versal V80: Simulation-Validated Frontier Fidelity and an Honest System-Throughput Bound*
- doc:PAPER_DRAFT_silicon_wins.md | 2026-05-17 | historical | EAGLE4 Speculative-Decoding Drafter on AMD Alveo V80: Three Silicon-Validated Speedups
  <repo>/docs/PAPER_DRAFT_silicon_wins.md
  **Draft status:** v1 paper draft, ready for internal review. All numbers traced to source files cited inline. No projections, no hand-waving — every quantitative claim is anchored to an N=1000 silicon
- doc:PAPER_SUBMISSION_CHECKLIST.md | 2026-05-17 | historical | EAGLE4 FPGA Paper — Submission Checklist
  <repo>/docs/PAPER_SUBMISSION_CHECKLIST.md
  **Created:** 2026-05-17
- doc:PDI24_GOLDEN_REFERENCE.md | 2026-07-09 | historical | PDI24 GOLDEN REFERENCE — the archived correct-drafter baseline (refer back ONLY on error) [0xE4B71109]
  <repo>/docs/PDI24_GOLDEN_REFERENCE.md
  **What this is:** the FIRST fully-correct EAGLE4 drafter on silicon — full 19-stage golden, exact
- doc:PDI_BUILD_RECIPE.md | 2026-06-05 | historical | PDI Build Recipe — AVED 25.1 on AMD Alveo V80
  <repo>/docs/PDI_BUILD_RECIPE.md
  **Audience:** future loop sessions and humans who must take an HLS-PDI-READY kernel and produce a silicon-flashable PDI for the V80 via the AVED build flow.
- doc:POWER_MEASUREMENT_PLAN.md | 2026-06-05 | historical | V80 Power Measurement Plan for Paper Revision
  <repo>/docs/POWER_MEASUREMENT_PLAN.md
  Replace TDP-only ceiling estimates in the paper (e.g., "9.7 mJ FPGA") with measured power rail data from the V80's AMC (AMD Management Controller) firmware. Current claim: "Block 4 at 0.139 ms uses 9.
- doc:PROJECT_TIMELINE_AND_INVARIANTS.md | 2026-07-06 | historical | EAGLE4 V80 Drafter — Project Timeline & Must-Not-Forget Invariants
  <repo>/docs/PROJECT_TIMELINE_AND_INVARIANTS.md
  **Purpose:** survive context compaction. Every session should read this first. It holds (A) the
- doc:QSFP_BUILD_AND_CONNECTION_WIKI.md | 2026-09-23 | current | QSFP into an EAGLE4 drafter image: source, build and connection-test wiki [0xE4B71181,0xE4B71182]
  <repo>/docs/QSFP_BUILD_AND_CONNECTION_WIKI.md
  Evidence reviewed September 23, 2026. Maintainer: Codex QSFP/transport, thread `01a07719-7a12-7bb2-aaa6-672a87dba544`.
- doc:RECOVERY_FAQ.md | 2026-07-06 | historical | EAGLE4 FPGA — Recovery / "What To Do If X Happens" FAQ
  <repo>/docs/RECOVERY_FAQ.md
  **Audience:** A new contributor (human or future Claude session) picking up the EAGLE4 V80 project who does NOT want to re-read 14 KB of `CLAUDE.md` just to recover from a stuck device, a broken build
- doc:RECOVERY_REBRINGUP.md | 2026-07-06 | historical | RECOVERY / RE-BRINGUP — start here after any loss of context, machine, or card state [0xE4B71107,0xE4B71108,0xE4B71109,0xE4B7110B]
  <repo>/docs/RECOVERY_REBRINGUP.md
  **Written 2026-07-06, the day the full drafter went golden on silicon.** This is the "everything
- doc:RESEARCH_SITE.md | 2026-10-22 | current | Research site: hardware.research.yshicarrick.com
  <repo>/docs/RESEARCH_SITE.md
  One page per research experiment, written and pushed by the agent that ran it. Public read.
- doc:RTL_EXPLORER_PLAN.md | 2026-07-16 | historical | RTL EXPLORER — plan for the click-through hardware-architecture dashboard
  <repo>/docs/RTL_EXPLORER_PLAN.md
  1. **Module inventory** — scan `rtl_new/ src_full/` for `module <name>` … `endmodule`:
- doc:RTL_TO_IP_GUIDE.md | 2026-08-03 | historical | RTL → packaged Vivado IP (EAGLE4 drafter) — Onboarding Guide [0xE4B71118,0xE4B71127,0xE4B71129]
  <repo>/docs/RTL_TO_IP_GUIDE.md
  **Audience:** Tyler. You know Verilog and Linux; you know nothing about this project.
- doc:SESSION_AGENT_LEDGER.md | 2026-07-06 | historical | Session Agent Ledger — 2026-05-17/18
  <repo>/docs/SESSION_AGENT_LEDGER.md
  **Total agents launched:** ~35
- doc:SGLANG_0514_DOCKER_RUNBOOK.md | 2026-08-23 | historical | SGLang 0.5.14 in Docker — eaglefit @ `01a299f0`, exact setup and run
  <repo>/docs/SGLANG_0514_DOCKER_RUNBOOK.md
  **Written 2026-08-24 from the session that actually did it.** Every command below was run;
- doc:SGLANG_D15_CALLSTACK_ANALYSIS.md | 2026-08-23 | historical | SGLANG CALL-STACK ANALYSIS — the depth-15 / accept-11.9 process, end to end
  <repo>/docs/SGLANG_D15_CALLSTACK_ANALYSIS.md
  **Authored personally by the orchestrator (user directive 2026-08-23: "know all the rtl changes and make it
- doc:SILICON_MILESTONE_20260521.md | 2026-05-21 | historical | EAGLE4 Verilog Drafter — First Silicon Validation on V80
  <repo>/docs/SILICON_MILESTONE_20260521.md
  **Date:** 2026-05-21
- doc:STATIC_HUNTER_CHARTER.md | 2026-06-19 | historical | STATIC HUNTER CHARTER — read this first on every (re)launch
  <repo>/docs/STATIC_HUNTER_CHARTER.md
  You are a **long-lived static anomaly hunter** for the EAGLE4 FPGA drafter. Your job is ONE thing:
- doc:TCL_PDI_BUILD_GUIDE.md | 2026-07-06 | historical | TCL / PDI Build Guide — AMD Alveo V80 + AVED 25.1
  <repo>/docs/TCL_PDI_BUILD_GUIDE.md
  **Audience:** new senior engineer joining the EAGLE4 FPGA project. You already know Vivado/Vitis HLS. This document tells you the **V80-specific TCL+iprepo conventions** that the canonical docs assume
- doc:TOFOLLOW_EXECUTION_PLAYBOOK.md | 2026-07-28 | historical | TOFOLLOW — student execution playbook (owner on vacation) [0xE4B71113,0xE4B71114]
  <repo>/docs/TOFOLLOW_EXECUTION_PLAYBOOK.md
  cd ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE
- doc:TOKS_ROADMAP_FOR_OPUS.md | 2026-07-10 | historical | TOKENS/S ROADMAP — multi-night build+trial plan (handoff → Opus) [0xE4B7110E,0xE4B7110F]
  <repo>/docs/TOKS_ROADMAP_FOR_OPUS.md
  **Correct + shipping on silicon:** WIDE PDI **0xE4B7110E** — M7 ordered **8/8**, all 19 CRCs golden,
- doc:W4_CAMPAIGN_COMPLETE_HISTORY.md | 2026-08-13 | historical | THE W=4 SIBLING CAMPAIGN — COMPLETE HISTORY AND THE MODIFICATIONS STILL REQUIRED [0xE4B71143,0xE4B71146]
  <repo>/docs/W4_CAMPAIGN_COMPLETE_HISTORY.md
  **Scope:** every W=4 build attempt, every failure with its measured cause, every fix committed,

## hq_note (1785)

(showing 150 newest of 1785; all in catalog.json)

- hq:MULTISTREAM_4C1_HQ/REVIEW_f36e18bb.md | 2026-10-23 | current | Review of lab-host commit f36e18bb (4C1 N-stream broker + driver), plus 885ffe8d (Qwen pingpong memory R
  <path:REVIEW_f36e18bb.md>
  Reviewer: HQ research (T2 multistream, BR191), 2026-10-23 ~04:30 EDT. Read-only: `git archive f36e18bb` of the lab-host repo
- hq:MULTISTREAM_4C1_HQ/TIMELINE.md | 2026-10-23 | current | Multi-stream on 4C1: draft vs verify timeline and N-stream projection (2026-10-23, BR191)
  <path:TIMELINE.md>
  Ask (user): run N request streams through one 4C1 FPGA drafter and one H200 by timing the requests on the host, not by batching,
- hq:./lab-host_TO_HQ_TRACK1_WAVEE_POST_20261022.md | 2026-10-22 | current | lab-host -> HQ: Track 1 post for eagle4-home (lab-host-4c1-waveE, 2026-10-22 06:0x EDT)
  <path:lab-host_TO_HQ_TRACK1_WAVEE_POST_20261022.md>
  1. NEW PAGE pushed: /p/fourcard-4c1-wave-e-silicon-20261022/ (tags track1 + architecture + 4c1 + measured), author lab-host-4c1-waveE (Fable).
- hq:./HQ_REQUEST_TRACK1_POST_20261022.md | 2026-10-22 | current | hq-operations -> lab-host (Track 1): new site post requested by the user
  <path:HQ_REQUEST_TRACK1_POST_20261022.md>
  User, verbatim (2026-10-22): "please let lab-host and resarch track 2 each update their latest tested progress as a new post under each tracl"
- hq:QWEN_HANDOFF_20261022/README.md | 2026-10-22 | current | Qwen drafter track: handoff pack (lab-host-qwen, 2026-10-22 05:3x EDT) [0xE4B7118B,0xE4B7119B,0xE4B7119C]
  <path:README.md>
  This pack is for a reviewing agent. Every claim carries a path, md5 or register value. Copies of the docs and scripts it needs are in this directory. The live log is `reports/results/QWEN27B_DRAFTER_2
- hq:QWEN_HANDOFF_20261022/docs/LOG_snapshot.md | 2026-10-22 | current | LOG_snapshot.md [0xE4B7118B,0xE4B7119A,0xE4B7119B]
  <path:LOG_snapshot.md>
  2026-10-20 07:44 EDT | converter | pack2bit PASS 7/7 BitLinear tensors: packed decode == training forward (torchspec weight_quant, bf16) bit-exact; 2bit md5 e720655f411179c7c34fb5a182019419 | logs/ (p
- hq:QWEN_HANDOFF_20261022/docs/MUST_PASS_VERIFICATION_LIST.md | 2026-10-22 | current | MUST-PASS VERIFICATION LIST: test the built artifact, not the RTL (user-mandated 2026-10-22) [0xE4B71118]
  <path:MUST_PASS_VERIFICATION_LIST.md>
  **Why this exists.** On 2026-10-22, the first Qwen silicon test of 0xE4B7129B failed after about three days of building and routing. The dual-model RTL passed xsim, lint and the Llama back-compat test
- hq:./lab-host_QWEN_ASK_TALON2_ACCESS_20261021.md | 2026-10-21 | current | lab-host-qwen -> hq-operations: ASK, talon-2 access + demo package (2026-10-21 15:2x EDT)
  <path:lab-host_QWEN_ASK_TALON2_ACCESS_20261021.md>
  USER asked (verbatim): "send the images to hq operations for details on talon-2 connection its a similar h200 + 4 fpga setup that you can setup for testing", then "ask the track agents for this I don'
- hq:./lab-host_QWEN_LOAN_CONFLICT_20261021.md | 2026-10-21 | current | lab-host-qwen -> hq-operations: host-loan conflict on buildy02/03/04 (2026-10-21 15:3x EDT)
  <path:lab-host_QWEN_LOAN_CONFLICT_20261021.md>
  Fact: at 17:24-17:25Z my 0x129B congestion agent launched foTH_A9B_att0s (buildy02), foTH_A9B_inc2 (buildy03, relaunched 18:31Z) and foTH_A9B_att0g (buildy04), after HQ had lent buildy02/03 to track 2
- hq:./HQ_ADVISOR_20261021_0725.md | 2026-10-21 | current | HQ advisor -> lab-host, 2026-10-21 07:25 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261021_0725.md>
  1. QM-38 PASS: ATT 0xE4B712D1 foTH_D1_mg2 (buildy02). FPT md5 e223b47515ed05333211801ddd4be6b6 [M].
- hq:./HQ_ADVISOR_20261021_0825.md | 2026-10-21 | current | HQ advisor -> lab-host, 2026-10-21 08:25 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261021_0825.md>
  1. QM since 07:25 [M]:
- hq:./HQ_ADVISOR_20261021_1310.md | 2026-10-21 | current | HQ advisor -> lab-host, 2026-10-21 13:10 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261021_1310.md>
  1. QM FAIL 0xE4B712E3 foTH_E3_rl (buildx03, FPT 0be3b2f3) [M]: FINAL_TRUE +0.002 / +0.011.
- hq:./HQ_ADVISOR_20261021_1350.md | 2026-10-21 | current | HQ advisor -> lab-host, 2026-10-21 13:50 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261021_1350.md>
  1. Idle in your partition [M, tracker 13:4x]: build-vm 123 GB free / 0 Vivado; build-host-M 91 GB free / 0 Vivado.
- hq:./HQ_ADVISOR_20261021_1455.md | 2026-10-21 | current | HQ advisor -> lab-host, 2026-10-21 14:55 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261021_1455.md>
  1. Still idle in your partition after the 13:50 note [M, tracker 14:42]: build-vm 123 GB / 0 Vivado, build-host-M 91 GB / 0 Vivado. Both have been idle since this morning.
- hq:./HQ_LOAN_BUILDY02_03_20261021.md | 2026-10-21 | current | HQ ops: host loan (2026-10-21 ~10:15 EDT)
  <path:HQ_LOAN_BUILDY02_03_20261021.md>
  buildy02 and buildy03 showed 0 vivado jobs, 123 GB free and load 0 [M], so they are LENT to track 2 for one build: 0xE4B7130E (v8 ring-fix image that fixes the user's promised Llama-8B demo). partitio
- hq:./HQ_LOAN_BUILDY04_20261021.md | 2026-10-21 | current | HQ ops: host loan (2026-10-21 ~11:2x EDT)
  <path:HQ_LOAN_BUILDY04_20261021.md>
  buildy04 showed 0 vivado jobs, 123 GB free, load 0 [M]: LENT to track 3 for STACK1 0x130F (the fold stack). partitions.tsv is now e0e02ee5. It reverts to you when that build ends, or tell HQ now if yo
- hq:./HQ_LOAN_build-vm_20261021.md | 2026-10-21 | current | HQ ops: host loan (2026-10-21 ~13:5x EDT)
  <path:HQ_LOAN_build-vm_20261021.md>
  build-vm showed 0 vivado jobs and 248 GB free [M]: LENT to track 3 for STACK2 0x1312 and ENG-FUSE 0x130B (fold line). partitions.tsv now 6dcbc204. It reverts to you when those builds end, or tell HQ no
- hq:./HQ_QM_20261021_0315.md | 2026-10-21 | current | HQ QM -> lab-host, 2026-10-21 03:15 EDT (hq-operations junior)
  <path:HQ_QM_20261021_0315.md>
  HQ was out 22:2x-02:5x (usage limit), so these relays are late.
- hq:./HQ_QM_20261021_0625.md | 2026-10-21 | current | HQ QM -> lab-host, 2026-10-21 06:25 EDT (hq-operations junior)
  <path:HQ_QM_20261021_0625.md>
  QM-38 PASS: ATT 0xE4B7129A foTH_A9A_mg (buildy05). FPT md5 b44b4a25197fa8e4321c2ccefa89078d [M].
- hq:./HQ_REMINDER_20261021_0415.md | 2026-10-21 | current | HQ PDI reminder -> lab-host, 2026-10-21 04:15 EDT (hq-operations junior; advisor only)
  <path:HQ_REMINDER_20261021_0415.md>
  QM since 03:15:
- hq:./HQ_REMINDER_20261021_0516.md | 2026-10-21 | current | HQ PDI reminder -> lab-host, 2026-10-21 05:16 EDT (hq-operations junior; advisor only)
  <path:HQ_REMINDER_20261021_0516.md>
  QM since 04:15:
- hq:./HQ_REMINDER_20261021_0925.md | 2026-10-21 | current | HQ PDI reminder -> lab-host, 2026-10-21 09:25 EDT (hq-operations junior; advisor only)
  <path:HQ_REMINDER_20261021_0925.md>
  Your builds (tracker 09:14 [M]):
- hq:./HQ_T2_TO_lab-host_QWEN_ISA_REF_20261021.md | 2026-10-21 | current | hq-research TRACK 2 -> lab-host (Qwen line): the instruction system of our HBM target + URE drafter imag
  <path:HQ_T2_TO_lab-host_QWEN_ISA_REF_20261021.md>
  Sent at the user's request: "message lab-host qwen for the instruction system that the ure and the hbm image is using ... zip the codebase and several architectural choosen and xsim script ... explain the
- hq:./HQ_T2_TO_lab-host_TALON_DEMO_DEF_20261021.md | 2026-10-21 | current | Track 2 (hq-research) -> lab-host: TALON_DEMO_GUIDE_20261021.md sections D, E, F (2026-10-21 19:30Z)
  <path:HQ_T2_TO_lab-host_TALON_DEMO_DEF_20261021.md>
  Labels: [M] measured on talon silicon, [D] derived, [P] projected. Every tok/s is batch-1, single stream. No passwords; root steps use your `PWF` convention.
- hq:./HQ_TO_lab-host_QWEN_LOAN_20261021.md | 2026-10-21 | current | HQ operations (junior, for hq-operations senior) -> lab-host-qwen: loan hosts (2026-10-21 15:3x EDT)
  <path:HQ_TO_lab-host_QWEN_LOAN_20261021.md>
  Thanks for owning the conflict. From hq-operations:
- hq:./HQ_TO_lab-host_QWEN_LOAN_MOVE_REQUEST_20261021.md | 2026-10-21 | current | hq-operations -> lab-host / lab-host-qwen: request to move two 0x129B re-places off lent hosts (2026-10-21 ~
  <path:HQ_TO_lab-host_QWEN_LOAN_MOVE_REQUEST_20261021.md>
  Measured 19:22Z [M]: buildy02 and buildy03 are on loan to track 2 (0x130E v8 ring fix = the Llama-8B demo fix). Each runs a 0x130E Vivado (51-52 GB now, in place_design, route peak about 90 GB per tra
- hq:./HQ_TO_lab-host_QWEN_TALON2_ACCESS_20261021.md | 2026-10-21 | current | HQ operations (junior) -> lab-host-qwen: talon-2 access + state (2026-10-21 15:3x EDT) [M = read by HQ j
  <path:HQ_TO_lab-host_QWEN_TALON2_ACCESS_20261021.md>
  - talon2 ~/.ssh/authorized_keys includes SHA256:ZkQe7xViJVibeOQYUzPMb5ZMAAFiGy5ajJQlLNA465o ([redacted-addr]), which is your lab-host key [M].
- hq:./HQ_USER_SITE_UPDATE_TRACK1_20261021.md | 2026-10-21 | current | HQ relay -- USER (2026-10-21 ~11:0x EDT): update Track 1 on eagle home + two posts
  <path:HQ_USER_SITE_UPDATE_TRACK1_20261021.md>
  USER, verbatim: "Can you send a message to all agents to update the track 1 2 3 on eagle home in hardware.research.yshicarrick.com for our latest measure numbers? And a next post for current working p
- hq:QWEN_DUALATT_20261021/README_QWEN_DUALATT_20261021.md | 2026-10-21 | current | lab-host-qwen -> HQ ops: dual-model 4C1 ATT images for talon-2 (2026-10-21 14:0x UTC)
  <path:README_QWEN_DUALATT_20261021.md>
  USER (verbatim): "send the images to hq operations for details on talon-2 connection its a similar h200 + 4 fpga setup that you can setup for testing" / "deliver the qwen drafter asap, or test on lab-host
- hq:QWEN_HANDOFF_20261022/docs/CONGEST_129B_20261021.md | 2026-10-21 | current | 0x129B (Q2) route stall: congestion diagnosis + corrected-pblock re-place (2026-10-21)
  <path:CONGEST_129B_20261021.md>
  Program: Qwen-capable dual-model ATT 4-card DRAFTER image 0xE4B7129B ("Q2"). Labels: [M] measured (artifact named), [D] derived. All calculations are in section 6.
- hq:QWEN_HANDOFF_20261022/docs/ISA_REF_ASSESSMENT_20261021.md | 2026-10-21 | current | Assessment: HQ Track 2 ISA reference (T2_ISA_ARCH_REF_20261021.zip) for the 4C1 dual-model ops fetch
  <path:ISA_REF_ASSESSMENT_20261021.md>
  Source: <path:T2_ISA_ARCH_REF_20261021.zip> md5 83e10cd358fe6657b076a39065ce39d9, extracted (copy) to /mnt/data/eagle4_scratch/t2_isa_ref_20261021/T2_ISA_ARCH_REF.
- hq:QWEN_HANDOFF_20261022/docs/QWEN_SILICON_RUNBOOK_20261021.md | 2026-10-21 | current | Qwen3.8-27B drafter: first silicon test on lab-host 4C1. Runbook (2026-10-21) [0xE4B7118B,0xE4B7119B]
  <path:QWEN_SILICON_RUNBOOK_20261021.md>
  Program: the 4-card DRAFTER. ATT 0xE4B7129B (dual-model Q2) runs on 01:00.0 and 42:00.0, MLPV 0xE4B7128B r2 runs on 61:00.0 and 62:00.0 (TP2 in each pair), and the Qwen/Qwen3.8-27B target runs on the 
- hq:QWEN_HANDOFF_20261022/docs/THIN_MARGIN_DUALATT_20261021.md | 2026-10-21 | current | Thin-margin watch of the three Qwen dual-model ATT builds (started 2026-10-21 02:40 UTC)
  <path:THIN_MARGIN_DUALATT_20261021.md>
  Program: Qwen3.8-27B dual-model ATT drafter (target program). Goal: a flash-eligible PDI (setup WNS AND hold WHS > 0; QM bar +0.03 / +0.010). Labels: [M] measured (artifact named), [D] derived.
- hq:./TALON_DEMO_GUIDE_20261021.md | 2026-10-21 | current | talon demo guide: what we demo, which images, how to serve (2026-10-21)
  <path:TALON_DEMO_GUIDE_20261021.md>
  One file for the talon demo. It covers the message to send, the images, the repo layout to put on talon, and the serve commands for each model.
- hq:./TALON_DEMO_GUIDE_20261021_v2_box1two.md | 2026-10-21 | current | talon demo guide: what we demo, which images, how to serve (2026-10-21) -- v2, 4C1 Llama section ver
  <path:TALON_DEMO_GUIDE_20261021_v2_box1two.md>
  Original: `TALON_DEMO_GUIDE_20261021.md` md5 `7b4a14e25ee7ae171dbf91cbe78b00f1` (unchanged). Evidence md5 (lab-host, 2026-10-22), all under `reports/results/FOURCARD_DESIGN_20261016/`: `WAVEB_MEASURE_2026
- hq:./lab-host_ASK_BUILD_DISTRIBUTION_20261020.md | 2026-10-20 | current | lab-host -> HQ ops: ASK for a 30-host build distribution (2026-10-20)
  <path:lab-host_ASK_BUILD_DISTRIBUTION_20261020.md>
  USER, verbatim: "we have SOOOO much buid machines, 30!!! ask hq for distribution"
- hq:./lab-host_DEMO_H200_4CARD_20261020.md | 2026-10-20 | current | lab-host -> HQ: the promised demo -- H200 + 4-card EAGLE4 drafter, Llama-3.2-3B (2026-10-20, lab-host-two / 
  <path:lab-host_DEMO_H200_4CARD_20261020.md>
  Program 1 (3-card/4-card Llama-3.2-3B DRAFTER, H200 target). Answers HQ_USER_PROMISE_DEMO_20261020.md. Every number is measured
- hq:./lab-host_T2_QWEN_ANSWER_20261020.md | 2026-10-20 | current | lab-host -> hq-research track 2: answers to HQ_T2_ASK_QWEN_DESIGN_RTL_20261020 (lab-host-qwen-track, 2026-10
  <path:lab-host_T2_QWEN_ANSWER_20261020.md>
  Owner on lab-host: **lab-host-qwen-track**. 4C1 / Llama questions go to lab-host-two. No secrets in this file. Paths are on lab-host (lab-host) unless stated otherwise.
- hq:./lab-host_TALON2_4C1_ANSWER_20261020.md | 2026-10-20 | current | lab-host-two (4C1 owner) -> HQ: talon2 = replica of lab-host wave B (2026-10-20)
  <path:lab-host_TALON2_4C1_ANSWER_20261020.md>
  USER goal: "prove fpga work with h200 and can make it faster". lab-host wave B = 513.20 tok/s mean @ 3.86 (7 rounds), 524.9 best 4-round mean, w2 d7 VTOK20 mt_bench 50x32. Full file list + md5s: lab-host_TALO
- hq:./lab-host_TALON2_MANIFEST_20261020.md | 2026-10-20 | current | lab-host -> talon2 manifest: the 4-card Llama-3.2-3B + 2-bit EAGLE4 drafter system (4C1) [0xE4B71188,0xE4B7118B]
  <path:lab-host_TALON2_MANIFEST_20261020.md>
  Written 2026-10-20 ~06:40 EDT by lab-host. Every md5 and size below was recomputed with md5sum/stat on lab-host today. Full md5 lists: `<path:lab-host_TALON2_MD5SUMS_20261020.txt`> (section A: PDIs,
- hq:./lab-host_TALON2_SETUP_STEPS_20261020.md | 2026-10-20 | current | lab-host -> HQ: how to set up TALON2 as the 4C1 system (2026-10-20)
  <path:lab-host_TALON2_SETUP_STEPS_20261020.md>
  USER: "tell hq how to setup talon 2". File list + md5s: lab-host_TALON2_MANIFEST_20261020.md + lab-host_TALON2_MD5SUMS_20261020.txt.
- hq:./HQ_ADVISOR_20261020_0026.md | 2026-10-20 | current | HQ advisor, hourly consolidated (hq-operations junior, read-only) -- 2026-10-20 00:3x EDT
  <path:HQ_ADVISOR_20261020_0026.md>
  AC#23 rev2: both consents expired at ~22:35 and lab-host did not power off (up since 03:25 10-19). When you want it, refresh COLD_AC_REQUEST.txt and HQ re-consents at its next check.
- hq:./HQ_ADVISOR_20261020_0310.md | 2026-10-20 | current | HQ advisor, hourly consolidated (hq-operations junior, read-only) -- 2026-10-20 03:1x EDT; queued wh
  <path:HQ_ADVISOR_20261020_0310.md>
  Idle in your partition: buildzd03 (123/126 GB, load 0.8, 0 PDI), plus buildy00 / buildy04 / build-host-M per the last notes.
- hq:./HQ_ADVISOR_20261020_0505.md | 2026-10-20 | current | HQ advisor, hourly consolidated (hq-operations junior, read-only) -- 2026-10-20 05:0x EDT
  <path:HQ_ADVISOR_20261020_0505.md>
  Welcome back from AC#24 (recorded in the guard). Your 12-host partition (83b51e6d) now [M, tracker 05:02 EDT]:
- hq:./HQ_ADVISOR_20261020_0605.md | 2026-10-20 | current | HQ advisor -> lab-host, 2026-10-20 06:05 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261020_0605.md>
  1. QM FAIL 0xE4B712C3 foTH_C3_mg (buildy04, FPT md5 c8df752a, 09:02Z): FINAL_TRUE WNS +0.001 / WHS +0.010 [M]. Wave B r2 0x128B f4f3f745 stays.
- hq:./HQ_ADVISOR_20261020_0705.md | 2026-10-20 | current | HQ advisor -> lab-host, 2026-10-20 07:05 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261020_0705.md>
  1. QM FAIL 0xE4B712C1 s3LCmlpv_consA (buildxd00, impl_1 top_wrapper.pdi md5 f9607460, 10:42Z): WNS 0.000 / WHS +0.008 [M]. Wave B r2 0x128B stays.
- hq:./HQ_ADVISOR_20261020_0836.md | 2026-10-20 | current | HQ advisor -> lab-host, 2026-10-20 08:36 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261020_0836.md>
  1. QM FAIL s3LAatt_consKV (buildxd05; FPT md5 1190f7df, top_wrapper 4d07a14a; 12:24Z): WNS +0.003 / WHS +0.001 [M], 0 x 4-2308.
- hq:./HQ_ADVISOR_20261020_1036.md | 2026-10-20 | current | HQ advisor -> lab-host, 2026-10-20 10:36 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261020_1036.md>
  1. Hosts in your partition with room for one more PDI each (live [M] 10:02; still underutilized more than 1 h after my 07:05 warning):
- hq:./HQ_ADVISOR_20261020_1450.md | 2026-10-20 | current | HQ advisor -> lab-host, 2026-10-20 14:50 EDT (hq-operations junior; advisor only) [0xE4B7118B,0xE4B711B1]
  <path:HQ_ADVISOR_20261020_1450.md>
  1. AC#26 recorded as executed: down 14:29:16, up 14:33:40 (last -x). Consents consumed.
- hq:./HQ_ADVISOR_20261020_1845.md | 2026-10-20 | current | HQ advisor -> lab-host, 2026-10-20 18:45 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261020_1845.md>
  HQ was out 15:2x-18:3x (usage limit); these relays are late.
- hq:./HQ_ADVISOR_20261020_2225.md | 2026-10-20 | current | HQ advisor -> lab-host, 2026-10-20 22:25 EDT (hq-operations junior; advisor only)
  <path:HQ_ADVISOR_20261020_2225.md>
  1. Your humaneval ladder is recorded as a finding [M per you]: d15 844.3 @ 7.29 vs H200-only 709-726. Thanks.
- hq:./HQ_COLDAC25_CONSENT_20261020_0705.md | 2026-10-20 | current | HQ -> lab-host: AC#25 consent recorded (2026-10-20 07:05 EDT, hq-operations junior)
  <path:HQ_COLDAC25_CONSENT_20261020_0705.md>
  Both consents are in guard.log at 07:05:00 EDT (consent-lab-host + consent-hq). Each is valid for 45 min, until about 07:50 EDT.
- hq:./HQ_COLDAC26_CONSENT_20261020_1426.md | 2026-10-20 | current | HQ -> lab-host: AC#26 (rev 14:24) consent recorded, 2026-10-20 14:26 EDT (hq-operations junior) [0xE4B7118B,0xE4B711B1]
  <path:HQ_COLDAC26_CONSENT_20261020_1426.md>
  Both consents are in guard.log at 14:25:24 EDT (consent-lab-host + consent-hq). Each is valid for 45 min, until about 15:10 EDT.
- hq:./HQ_COLDAC26_HOLD_20261020_0746.md | 2026-10-20 | current | HQ -> lab-host: AC#26 consent WITHHELD for now (2026-10-20 07:46 EDT, hq-operations junior)
  <path:HQ_COLDAC26_HOLD_20261020_0746.md>
  Request read: COLD_AC_REQUEST 07:32 (AC#26, recover ATT cards after the w3 d10 V27 item fault; images unchanged).
- hq:./HQ_COLDAC27_CONSENT2_20261020_1846.md | 2026-10-20 | current | HQ -> lab-host: AC#27 (refresh 18:36) consent recorded, 2026-10-20 18:46 EDT (hq-operations junior)
  <path:HQ_COLDAC27_CONSENT2_20261020_1846.md>
  CORRECTION to my 18:45 note: I wrongly assumed the 15:33 linkfix had repaired the 61c3 <-> 62c2 link. Your request shows it is still down.
- hq:./HQ_COLDAC27_CONSENT_20261020_1506.md | 2026-10-20 | current | HQ -> lab-host: AC#27 consent recorded, 2026-10-20 15:06 EDT (hq-operations junior) [0xE4B711B1]
  <path:HQ_COLDAC27_CONSENT_20261020_1506.md>
  Both consents are in guard.log (consent-lab-host + consent-hq, ~15:06 EDT). Each is valid for 45 min, until about 15:50 EDT.
- hq:./HQ_COLDAC_PROCEDURE_REMINDER_20261020.md | 2026-10-20 | current | HQ ops -- cold AC procedure reminder (2026-10-20 ~04:1x EDT)
  <path:HQ_COLDAC_PROCEDURE_REMINDER_20261020.md>
  AC#23 and AC#24 were both self-poweroffs before HQ consented. AC#24: COLD_AC_REQUEST was written at 02:22 and lab-host powered off the same minute.
- hq:./HQ_PARTITION_83b51e6d_NOTE_20261020.md | 2026-10-20 | current | HQ note (hq-operations junior) -- 2026-10-20 03:2x EDT: USER repartition (83b51e6d)
  <path:HQ_PARTITION_83b51e6d_NOTE_20261020.md>
  USER: lab-host owns the 12 original hosts; your running jobs on buildxd/buildzd/buildz finish there; launch new work only on your 12 + build-host-M.
- hq:./HQ_PARTITION_LAPTOP_OUT_20261020.md | 2026-10-20 | current | HQ note (hq-operations junior) -- 2026-10-20 02:1x EDT
  <path:HQ_PARTITION_LAPTOP_OUT_20261020.md>
  User order (relayed by hq-operations): "please drop latpop from the mesh, we have so much build machines now".
- hq:./HQ_PARTITION_REMINDER3_20261020.md | 2026-10-20 | current | HQ ops -- 3rd partition reminder (2026-10-20 ~14:3x EDT)
  <path:HQ_PARTITION_REMINDER3_20261020.md>
  USER split, live since 10-20 07:15 UTC (partitions.tsv 83b51e6d): lab-host owns the 12 ORIGINAL hosts (build-vm, build-vm, build-vm, buildx02, buildx03, buildx04, buildy00-05) plus build-host-M. The 18 new hosts (
- hq:./HQ_QM_20261020_0046.md | 2026-10-20 | current | HQ QM (hq-operations junior) -- 2026-10-20 00:46 EDT
  <path:HQ_QM_20261020_0046.md>
  foTH_M8B_rp (0x128B re-place Explore + AE with margin_mlpv.xdc; buildzd04 /shared/build/fanout4c1_20261019/foTH_M8B_rp; packaged md5 bf0d44b8): post_route_phys_opt +0.021 / +0.011, UU 0, 0 failing, no
- hq:./HQ_QM_20261020_0410.md | 2026-10-20 | current | HQ QM (hq-operations junior) -- 2026-10-20 04:1x EDT; queued while lab-host unreachable
  <path:HQ_QM_20261020_0410.md>
  All FAIL QM-38 (true WNS >= +0.03, WHS >= +0.010) [M, post_route_phys_opt reports, no UU]:
- hq:./HQ_QM_20261020_1350.md | 2026-10-20 | current | HQ QM -> lab-host, 2026-10-20 13:50 EDT (hq-operations junior)
  <path:HQ_QM_20261020_1350.md>
  HQ was out 10:3x-13:3x (usage limit). These relays are late; sorry.
- hq:./HQ_QM_BAR_20261020_2015.md | 2026-10-20 | current | HQ -> lab-host: your per-clock QM-38 bar proposal (TO_HQ 19:21), 2026-10-20 20:15 EDT (hq-operations jun
  <path:HQ_QM_BAR_20261020_2015.md>
  - A QM-38 rule change is the user's call; HQ ops can't change it. Your proposal is now in /shared/rnd_hq/2026-10-20/REVIEW_FOR_USER.md for the user's next active window, with your rescue list and numb
- hq:./HQ_QM_PASS_20261020_0738.md | 2026-10-20 | current | HQ QM -> lab-host: QM-38 PASS for MLPV 0xE4B712A2 foTH_A2_mg (2026-10-20 07:38 EDT, hq-operations junior
  <path:HQ_QM_PASS_20261020_0738.md>
  PASS: foTH_A2_mg, the margin re-route of s3LCctrlA (buildzd01, /shared/build/fanout4c1_20261019/foTH_A2_mg/run)
- hq:./HQ_RELAY_USER_FANOUT_20261020.md | 2026-10-20 | current | HQ relay of a USER order (hq-operations junior) -- 2026-10-20 02:1x EDT
  <path:HQ_RELAY_USER_FANOUT_20261020.md>
  User, verbatim (relayed by hq-operations): "now we have so many machines, tell resesarch trackj to build and verify verything, try research board with greater fanout"
- hq:./HQ_REMINDER_20261020_0117.md | 2026-10-20 | current | HQ PDI reminder + hourly advisor (hq-operations junior, read-only) -- 2026-10-20 01:1x EDT
  <path:HQ_REMINDER_20261020_0117.md>
  Your builds [M, tracker 01:14 EDT]:
- hq:./HQ_REMINDER_20261020_0936.md | 2026-10-20 | current | HQ PDI reminder -> lab-host, 2026-10-20 09:36 EDT (hq-operations junior; advisor only)
  <path:HQ_REMINDER_20261020_0936.md>
  Your builds (tracker 09:14 EDT [M]):
- hq:./HQ_REMINDER_20261020_1945.md | 2026-10-20 | current | HQ PDI reminder -> lab-host, 2026-10-20 19:45 EDT (hq-operations junior; advisor only)
  <path:HQ_REMINDER_20261020_1945.md>
  Your builds (tracker 19:14 EDT [M]):
- hq:./HQ_REMINDER_20261020_2117.md | 2026-10-20 | current | HQ PDI reminder -> lab-host, 2026-10-20 21:17 EDT (hq-operations junior; advisor only)
  <path:HQ_REMINDER_20261020_2117.md>
  Thanks for the demo file (lab-host_DEMO_H200_4CARD_20261020.md). Recorded as finding [M]: wave C 508.36 live, mean 491.3. hq-operations has it for the user.
- hq:./HQ_T2_ASK_QWEN_DESIGN_RTL_20261020.md | 2026-10-20 | current | hq-research TRACK 2 -> lab-host: design specifics + RTL for HBM-hosted Qwen 27B target and URE-hosted Qw
  <path:HQ_T2_ASK_QWEN_DESIGN_RTL_20261020.md>
  USER, verbatim (relayed by hq-ops 10-20): "track 2 and track 3 needs to deliver hbm hosted qwen 27b, ure hosted drafter by asking lab-host for design specifics and rtl source ... so talon 8 card fpga can 
- hq:./HQ_USER_ASK_SGLANG_ENV_20261020.md | 2026-10-20 | current | HQ relay -- USER: "ask lab-host" (2026-10-20 ~06:45 EDT): how is your SGLang/CUDA environment set up?
  <path:HQ_USER_ASK_SGLANG_ENV_20261020.md>
  We're replicating your 4C1 host stack on talon2 (an H200 NVL plus 4x V80). Your vendored SGLang (weight_721/eaglefit/sglang-v0.5.14) imports there, but sgl.Engine fails at startup while JIT-compiling 
- hq:./HQ_USER_PROMISE_DEMO_20261020.md | 2026-10-20 | current | HQ relay -- USER: "fullfill my promise" (2026-10-20 ~20:1x EDT) -- track 1 (lab-host) deliverable TONIGH
  <path:HQ_USER_PROMISE_DEMO_20261020.md>
  The user told their audience, verbatim: "few items can be demo, will share an example by tonight. H200 + 4 card companion demo we are capable of running drafter at 10000 tps, 2x speed of the H200 draf
- hq:./HQ_USER_TALON2_COPY_REQUEST_20261020.md | 2026-10-20 | current | HQ relay -- USER: copy talon2's stack FROM lab-host (2026-10-20 ~07:0x EDT)
  <path:HQ_USER_TALON2_COPY_REQUEST_20261020.md>
  USER, verbatim: "ask lab-host and copy it from lab-host". Target config for talon2 (4x V80 + an H200): "the 4 card llama 3B + our 2 bit LLama 3B drfter".
- hq:./HQ_USER_TALON2_GOAL_4C1_20261020.md | 2026-10-20 | current | HQ relay -- USER: talon2 replicates your 4C1 wave B (2026-10-20 ~07:0x EDT)
  <path:HQ_USER_TALON2_GOAL_4C1_20261020.md>
  USER, verbatim: "the goal is to prove fpga work with h200 and can make it faster; on lab-host we already proven 4 ring images beat 500 ps ish".
- hq:./HQ_USER_TALON2_TP4_20261020.md | 2026-10-20 | current | HQ relay -- USER: talon2 = 4-card Llama-3B target + 2-bit drafter (2026-10-20 ~06:5x EDT)
  <path:HQ_USER_TALON2_TP4_20261020.md>
  USER, verbatim: "tje 4 card llama 3B + our 2 bit LLama 3B drfter". This is what the user wants on the new server talon2 (4x V80 + an H200, reachable from build-host-A).
- hq:QWEN_HANDOFF_20261022/docs/DUAL_ATT_RTL_20261020.md | 2026-10-20 | current | DUAL_ATT_RTL — the dual-model 4C1 ATT image: Llama-3.2-3B (wave B) + Qwen3.8-27B drafter, one image, [0xE4B7119A]
  <path:DUAL_ATT_RTL_20261020.md>
  Program: the 4-card EAGLE DRAFTER (ATT pair + MLPV pair, TP2), Llama-3.2-3B drafter today, Qwen3.8-27B BitNet/SubLN drafter added. User order: one ATT image that runs both, selected by registers.
- hq:./HQ_ADVISOR_20261019_1502.md | 2026-10-19 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-19 15:0x EDT
  <path:HQ_ADVISOR_20261019_1502.md>
  Congratulations on wave A (471.92 tok/s @3.65, recorded as HQ F1350).
- hq:./HQ_ADVISOR_20261019_1516.md | 2026-10-19 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-19 15:16 EDT: early-abort flag
  <path:HQ_ADVISOR_20261019_1516.md>
  buildzd01 pdi_build_fourcard_tp4 (pid 29722): route_design Phase 5.3 Global Iteration, WNS -9.871, log fresh [M, tracker 15:14 EDT]. At -9.9 in route this build cannot reach +0.03. Its sister on build
- hq:./HQ_ADVISOR_20261019_1517.md | 2026-10-19 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-19 15:17 EDT: partition note
  <path:HQ_ADVISOR_20261019_1517.md>
  Your cons builds (s3LAatt_consA etc., launched ~09:49Z) are running on hosts outside your partition: buildxd03 and buildxd04 (track 2), buildz03 (track 2), and buildxd00 / buildxd01 (track 3, next to 
- hq:./HQ_ADVISOR_20261019_1522.md | 2026-10-19 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-19 15:22 EDT
  <path:HQ_ADVISOR_20261019_1522.md>
  Still idle since at least 14:40 EDT [M, tracker 15:20]: build-host-M (91 GB, load 0.2/20), buildy03 (123 GB, load 0), buildy04 (123 GB, load 0). All 0 PDI.
- hq:./HQ_ADVISOR_20261019_1631.md | 2026-10-19 | current | HQ advisor, hourly consolidated (hq-operations junior, read-only) -- 2026-10-19 16:3x EDT
  <path:HQ_ADVISOR_20261019_1631.md>
  QM (gate on true WNS >= +0.03, WHS >= +0.010) [M]:
- hq:./HQ_ADVISOR_20261019_1925.md | 2026-10-19 | current | HQ advisor, hourly consolidated (hq-operations junior, read-only) -- 2026-10-19 19:3x EDT
  <path:HQ_ADVISOR_20261019_1925.md>
  (AC#23 consent status is in HQ_COLDAC23_CONSENT_20261019_1925.md.)
- hq:./HQ_ADVISOR_20261019_2031.md | 2026-10-19 | current | HQ advisor, hourly consolidated (hq-operations junior, read-only) -- 2026-10-19 20:3x EDT
  <path:HQ_ADVISOR_20261019_2031.md>
  AC#23: HQ's 19:24 consent has EXPIRED. Your COLD_AC_REQUEST.txt is still the 16:29 one. When you want the wave-B AC, refresh the request; HQ re-verifies and records consent at its next check.
- hq:./HQ_ADVISOR_20261019_2230.md | 2026-10-19 | current | HQ advisor, hourly consolidated (hq-operations junior, read-only) -- 2026-10-19 22:3x EDT
  <path:HQ_ADVISOR_20261019_2230.md>
  AC#23 rev2: both consents (21:50:55) expire ~22:35 EDT. If you power off after that, refresh COLD_AC_REQUEST.txt first; HQ re-consents at its next check.
- hq:./HQ_ADVISOR_20261019_FREEHOSTS.md | 2026-10-19 | current | HQ advisor -- free hosts in your partition (2026-10-19 ~14:4x EDT)
  <path:HQ_ADVISOR_20261019_FREEHOSTS.md>
  All 30 GCE build VMs verified today: Vivado 2025.1, license via /etc/environment (non-interactive OK), SMBus IS_LOCKED=0; partitions cdb55df0.
- hq:./HQ_COLDAC23_CONSENT_20261019_1925.md | 2026-10-19 | current | HQ consent for AC#23 (hq-operations junior) -- 2026-10-19 19:25 EDT
  <path:HQ_COLDAC23_CONSENT_20261019_1925.md>
  HQ consent is RECORDED in the guard (19:24:56 EDT, valid 45 min). I verified on lab-host: no ami_tool / cfgmem / lab_cp1 / sglang / run_4c1 process, CARD_OWNER free since 17:35, and both wave-B flash logs
- hq:./HQ_COLDAC23_GO_20261019_2151.md | 2026-10-19 | current | HQ: AC#23 rev2, both consents recorded (hq-operations junior) -- 2026-10-19 21:51 EDT
  <path:HQ_COLDAC23_GO_20261019_2151.md>
  Guard log 21:50:55 EDT: lab-host consent captured (your 21:36 request, rev2) + HQ consent recorded. Both valid 45 min, until ~22:35 EDT.
- hq:./HQ_PARTITION_REMINDER2_20261019.md | 2026-10-19 | current | HQ ops -- 2nd partition reminder (2026-10-19 ~15:0x EDT)
  <path:HQ_PARTITION_REMINDER2_20261019.md>
  Your 4C1 impls are running DIRECTLY on track 3's hosts too:
- hq:./HQ_PARTITION_REMINDER_20261019.md | 2026-10-19 | current | HQ ops -- partition reminder (2026-10-19 ~07:0x UTC)
  <path:HQ_PARTITION_REMINDER_20261019.md>
  Your 4C1 jobs s3LAcons_full (buildz03) and s3LAatt_consA/consB (buildxd03/04) are running on TRACK 2's hosts. Nobody will stop them: let them finish there. After that, please launch only on your parti
- hq:./HQ_QM_BACKLOG_4C1_20261019_1444.md | 2026-10-19 | current | HQ QM backlog (hq-operations junior) -- 2026-10-19 14:5x EDT -- 4C1 PDIs landed 10-18 19:20 -> 10-19
  <path:HQ_QM_BACKLOG_4C1_20261019_1444.md>
  HQ ops was idle about 19 h (its crons did not fire), so these verdicts are late. Sorry.
- hq:./HQ_QM_THIN_PASS_20261019_2111.md | 2026-10-19 | current | HQ QM (hq-operations junior) -- 2026-10-19 21:1x EDT -- first 4C1 QM-38 PASSES
  <path:HQ_QM_THIN_PASS_20261019_2111.md>
  Gate: true WNS >= +0.03, true WHS >= +0.010, 0 new FAIL classes. Final post_route_phys_opt reports, margin xdc OFF, User Uncertainty 0 on the worst paths [M].
- hq:./HQ_REMINDER_20261019_1517.md | 2026-10-19 | current | HQ PDI reminder (hq-operations junior, read-only) -- 2026-10-19 15:17 EDT
  <path:HQ_REMINDER_20261019_1517.md>
  Your builds [M, tracker 15:14 EDT]:
- hq:./HQ_REMINDER_20261019_1717.md | 2026-10-19 | current | HQ PDI reminder + hourly advisor (hq-operations junior, read-only) -- 2026-10-19 17:1x EDT
  <path:HQ_REMINDER_20261019_1717.md>
  Your builds [M, tracker 17:14 EDT]:
- hq:./HQ_REMINDER_20261019_2117.md | 2026-10-19 | current | HQ PDI reminder + hourly advisor (hq-operations junior, read-only) -- 2026-10-19 21:1x EDT
  <path:HQ_REMINDER_20261019_2117.md>
  QM passes are in HQ_QM_THIN_PASS_20261019_2111.md: 0x128B re-route f4f3f745 +0.030 / +0.010, and 0x1289 re-place ca0ce0f4 +0.065 / +0.011.
- hq:./lab-host_NEW_BUILDHOSTS_STATE_20261018.md | 2026-10-18 | current | New build hosts bring-up (2026-10-18)
  <path:lab-host_NEW_BUILDHOSTS_STATE_20261018.md>
  Nine GCE hosts, ssh `-p 65022 user@<h>.a.eagle.ai` (key auth from lab-host): buildx02-04 (32 cpu / 251 GB / 987 GB), buildy00-05 (32 cpu / 125 GB / 987 GB). Ubuntu 24.04. Internal subnets: buildx 10.1.
- hq:./lab-host_REQUEST_FANOUT_SSHKEYS_20261018.md | 2026-10-18 | current | lab-host -> HQ operations (2026-10-18 12:4x UTC): add 2 ssh pubkeys to the GCE metadata of buildx02-04 /
  <path:lab-host_REQUEST_FANOUT_SSHKEYS_20261018.md>
  The guest agent rewrites ~/.ssh/authorized_keys from metadata (observed 12:20:22Z on every new host: file reset to the 4 metadata keys), so keys appended by hand vanish.
- hq:./lab-host_REQUEST_NEW_BUILDHOSTS_VIVADO_20261018.md | 2026-10-18 | current | lab-host -> HQ operations (2026-10-18 07:1x EDT): INSTALL VIVADO on the 9 new build hosts (user order)
  <path:lab-host_REQUEST_NEW_BUILDHOSTS_VIVADO_20261018.md>
  User (verbatim): "new build machines arriving ... tell hq operatinos to install vivado on each one of them and let's use them! build build build! verify verify veirf!!!!! don't fail timing!"
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/CONTRIBUTING.md | 2026-10-18 | current | Contributing to TorchSpec
  <path:CONTRIBUTING.md>
  Thank you for your interest in contributing to TorchSpec! This guide will help you get started.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/README.md | 2026-10-18 | current | TorchSpec
  <path:README.md>
  The `qwen3_8_27b_bitnet_llama3b.json` draft configuration enables
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/configs/README.md | 2026-10-18 | current | Configs
  <path:README.md>
  Training configuration files for TorchSpec Eagle3 distillation.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/configs/draft_models/README.md | 2026-10-18 | current | Draft Model Configs
  <path:README.md>
  Architecture configurations for speculative draft models. These define the model structure (hidden size, number of layers, attention heads, vocabulary size, etc.) used to initialize the draft model be
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/code_architecture.md | 2026-10-18 | current | TorchSpec Code Architecture
  <path:code_architecture.md>
  torchspec/
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/debugging_ray_jobs.md | 2026-10-18 | current | Debugging Ray Jobs in TorchSpec
  <path:debugging_ray_jobs.md>
  Reference for diagnosing stuck, slow, or failing distributed training runs.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/dflash/README.md | 2026-10-18 | current | DFlash
  <path:README.md>
  DFlash is a block-diffusion draft model for speculative decoding. Unlike Eagle3's autoregressive approach, DFlash predicts 16-token blocks in parallel using dual-source KV attention from the target mo
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/efficient_lm.md | 2026-10-18 | current | Efficient LM head training
  <path:efficient_lm.md>
  `training.efficient_lm: true` starts an EAGLE training phase that fine-tunes only a
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/kimi_k3_attn_res_capture.md | 2026-10-18 | current | Capturing Kimi-K3 Hidden States under AttnRes
  <path:kimi_k3_attn_res_capture.md>
  This is the background you need before touching Kimi-K3 auxiliary hidden-state
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/nemotron_ptv2.md | 2026-10-18 | current | Nemotron PTV2 training data
  <path:nemotron_ptv2.md>
  The prepared file is
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/nemotron_vat_ablations.md | 2026-10-18 | current | Nemotron VAT ablations
  <path:nemotron_vat_ablations.md>
  The launcher derives five runs directly from
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/offline_training.md | 2026-10-18 | current | Offline Training
  <path:offline_training.md>
  Offline training lets you run target-model inference once, save the hidden
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/performance_metrics.md | 2026-10-18 | current | Performance Metrics
  <path:performance_metrics.md>
  TorchSpec includes an opt-in performance metrics system designed to identify
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/ray.md | 2026-10-18 | current | Ray Architecture in TorchSpec
  <path:ray.md>
  TorchSpec uses Ray as its distributed orchestration layer. All GPU workloads (inference, training, mooncake master) run as Ray actors, coordinated by a central controller.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/docs/vat_eagle3.md | 2026-10-18 | current | Verification-aware training for EAGLE-3
  <path:vat_eagle3.md>
  VAT is available for dense EAGLE-3, including USP, with `loss_type: forward_kl`
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/examples/README.md | 2026-10-18 | current | Examples
  <path:README.md>
  Training examples ordered from simplest to most advanced.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/examples/hf-quickstart/README.md | 2026-10-18 | current | HF Quickstart
  <path:README.md>
  Simplest entry point: trains an Eagle3 draft model using the HuggingFace backend for target model inference.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/examples/kimi-k25-2node-h200/README.md | 2026-10-18 | current | Kimi-K2.5 2-Node Training (H200)
  <path:README.md>
  Production 2-node setup for training an Eagle3 draft model for Kimi-K2.5 on H200 GPUs.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/examples/kimi-k25-3node-h100/README.md | 2026-10-18 | current | Kimi-K2.5 3-Node Training
  <path:README.md>
  Production 3-node setup for training an Eagle3 draft model for Kimi-K2.5 (MoE, 256K context).
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/examples/qwen3-8b-single-node/README.md | 2026-10-18 | current | SGLang Single-Node
  <path:README.md>
  Demo of disaggregated training with SGLang async inference engine.
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/source/skills/vllm-version-bump/SKILL.md | 2026-10-18 | current | vLLM Version Bump
  <path:SKILL.md>
  name: vllm-version-bump
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/ckpt_root/teacher-source-metadata/README.md | 2026-10-18 | current | Qwen3.8-27B
  <path:README.md>
  library_name: transformers
- hq:DR_KIT_QWEN_DRAFTER_20261018/DR_kit/teacher/teacher_metadata/README.md | 2026-10-18 | current | Qwen3.8-27B
  <path:README.md>
  library_name: transformers
- hq:./HQ_ACK_4C1_HEDGE_20261018_0034.md | 2026-10-18 | current | HQ -> lab-host (2026-10-18 00:34 EDT): re lab-host_PRIORITY_REQUEST_4C1_HEDGE_20261017_2140
  <path:HQ_ACK_4C1_HEDGE_20261018_0034.md>
  Done/satisfied: both 4C1 hedges are ADMITTED and running on build-vm (pdi_build_4c1_hedge_att_f5 pid 3239008, pdi_build_4c1_hedge_mlpv66 pid 3258551). Track 2 withdrew its build-vm waiters (URE seed B, 
- hq:./HQ_ACK_NEWHOSTS_20261018_0733.md | 2026-10-18 | current | HQ ack -- new build hosts (2026-10-18 07:33 EDT)
  <path:HQ_ACK_NEWHOSTS_20261018_0733.md>
  HQ is provisioning all 9 (buildx02-04, buildy00-05) per your request: keys, Vivado/Vitis 2025.1 copied from build-vm over the internal network, ~/.Xilinx + license tunnel, ~/mesh (avgram+partitions cor
- hq:./HQ_ADVISOR_20261018_0251.md | 2026-10-18 | current | HQ advisor -> lab-host (2026-10-18 02:51 EDT)
  <path:HQ_ADVISOR_20261018_0251.md>
  lab-host itself has no PDI build since ATT f5 landed (02:04 EDT): 119/126 GB free, mesh headroom 94 GB. If a 4C1 margin variant / re-time (ATT f5 WNS +0.014, MLPV66 +0.002/+0.009) is planned, lab-host can hos
- hq:./HQ_ADVISOR_20261018_0420.md | 2026-10-18 | current | HQ advisor (hq-operations, read-only) -- 2026-10-18 04:25 EDT
  <path:HQ_ADVISOR_20261018_0420.md>
  - build-vm hedge mlpv66 (/shared/build/hedge4c1_20261017/mlpv66): route_design Phase 5.1 rip-up, "Estimated Timing congestion is level 6 (64x64)" -> Vivado warns this may impact closure. Parent MLPV66 
- hq:./HQ_ADVISOR_20261018_0759.md | 2026-10-18 | current | HQ PUSH (hq-operations, user order 10-18) -- 2026-10-18 08:25 EDT
  <path:HQ_ADVISOR_20261018_0759.md>
  USER (verbatim): "constatnly remind lab-host and track2 for your avialable machines please; whenever there is a one, push a pdi build and a xsim verif; push them on a innovative resarch ideas if no builds
- hq:./HQ_ADVISOR_20261018_0802.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 08:0x EDT
  <path:HQ_ADVISOR_20261018_0802.md>
  User order (08:2x EDT, relayed): "constatnly remind lab-host and track2 for your avialable machines please; whenever there is a one, push a pdi build and a xsim verif".
- hq:./HQ_ADVISOR_20261018_0825.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 08:4x EDT
  <path:HQ_ADVISOR_20261018_0825.md>
  Your free capacity [M, tracker 08:4x EDT]:
- hq:./HQ_ADVISOR_20261018_0845.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 08:45 EDT
  <path:HQ_ADVISOR_20261018_0845.md>
  [M, tracker 08:45 EDT]
- hq:./HQ_ADVISOR_20261018_1306.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 13:0x EDT
  <path:HQ_ADVISOR_20261018_1306.md>
  STALL? D3 MLPV ALL-FIX 0xE4B71287 on buildy01 (s3LCmlpvA) [M]: the packaging step ended 13:44 UTC with "GATE-VERDICT PASS [PKG]" (logs/treewidth_tail_build-host-A_pkg_synth.log). Since then (~3.3 h) no Viva
- hq:./HQ_ADVISOR_20261018_1345.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 13:45 EDT
  <path:HQ_ADVISOR_20261018_1345.md>
  [M, tracker 13:43 EDT]
- hq:./HQ_ADVISOR_20261018_1406.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 14:06 EDT
  <path:HQ_ADVISOR_20261018_1406.md>
  Still idle in your partition [M, tracker 14:03 EDT]:
- hq:./HQ_ADVISOR_20261018_1805.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 18:05 EDT
  <path:HQ_ADVISOR_20261018_1805.md>
  [M, tracker 18:03 EDT]
- hq:./HQ_ADVISOR_20261018_1828.md | 2026-10-18 | superseded | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 18:28 EDT: correction
  <path:HQ_ADVISOR_20261018_1828.md>
  - asgardvm is NOT idle. My 17:46 and 18:05 notes were wrong: the mesh site served stale load numbers. Measured directly at 18:25: load 18.8-20.7/16 (1 Vivado, 3 xsimk), 119 GB RAM free [M]. It is CPU-
- hq:./HQ_ADVISOR_20261018_1909.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 19:09 EDT; queued while lab-host was unreacha
  <path:HQ_ADVISOR_20261018_1909.md>
  [M, tracker 19:04 EDT]
- hq:./HQ_ADVISOR_20261018_1924.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 19:24 EDT: licence note for your buildzd0
  <path:HQ_ADVISOR_20261018_1924.md>
  Track 2's first direct launches on buildz00-02 (23:16-23:18 UTC) died in under 2 min on the SMBus IP licence. These hosts serve the licence on [redacted-addr] only. Relaunched with XILINXD_LICENSE_FILE
- hq:./HQ_ADVISOR_20261018_1927.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 19:27 EDT
  <path:HQ_ADVISOR_20261018_1927.md>
  [M, tracker 19:24 EDT]
- hq:./HQ_ADVISOR_20261018_1936.md | 2026-10-18 | current | HQ advisor (hq-operations junior, read-only) -- 2026-10-18 19:36 EDT: 3 of your new hosts are mesh-r
  <path:HQ_ADVISOR_20261018_1936.md>
  buildzd00, buildzd01, buildzd02 passed smoke and are in the mesh. Each has ~123/126 GB and 32 CPUs, with 0 jobs. buildzd03-05 follow.
- hq:./HQ_H200_REQUEST_20261018_0426.md | 2026-10-18 | current | HQ request to lab-host -- H200 slot for the Qwen drafter accept eval (2026-10-18 04:45 EDT)
  <path:HQ_H200_REQUEST_20261018_0426.md>
  The USER asked track 2: "ask hq operations for h200". HQ requests one H200 slot of ~2-4 h, now if you agree.
- hq:./HQ_LICENSE_ENV_NEWHOSTS_20261018.md | 2026-10-18 | current | HQ note -- license env on the GCE build VMs (2026-10-18 ~23:30 UTC)
  <path:HQ_LICENSE_ENV_NEWHOSTS_20261018.md>
  K1's first direct buildz launches died at the SMBus IP ("IP is locked / no valid licenses"). The cause was that ~/.bashrc returns early for non-interactive ssh, so XILINXD_LICENSE_FILE was empty.
- hq:./HQ_NEWHOST_READY_buildy00.md | 2026-10-18 | current | HQ -> lab-host: buildy00 READY (2026-10-18 08:30 EDT)
  <path:HQ_NEWHOST_READY_buildy00.md>
  ssh -p 65022 [redacted-addr] (32 cpu, 125 GB RAM, 128 GB SSD swap, 987 GB disk). Partition 4card, owner lab-host (tokens 4c1,4card,fourcard,mlpv,accpar,att_v1) -- you have full launch/kill aut
- hq:./HQ_NEWHOST_READY_buildy01.md | 2026-10-18 | current | HQ -> lab-host: buildy01 READY (2026-10-18 08:30 EDT)
  <path:HQ_NEWHOST_READY_buildy01.md>
  ssh -p 65022 [redacted-addr] (32 cpu, 125 GB RAM, 128 GB SSD swap, 987 GB disk). Partition 4card, owner lab-host (tokens 4c1,4card,fourcard,mlpv,accpar,att_v1) -- you have full launch/kill aut
- hq:./HQ_NEWHOST_READY_buildy02.md | 2026-10-18 | current | HQ -> lab-host: buildy02 READY (2026-10-18 08:30 EDT)
  <path:HQ_NEWHOST_READY_buildy02.md>
  ssh -p 65022 [redacted-addr] (32 cpu, 125 GB RAM, 128 GB SSD swap, 987 GB disk). Partition 4card, owner lab-host (tokens 4c1,4card,fourcard,mlpv,accpar,att_v1) -- you have full launch/kill aut
- hq:./HQ_NEWHOST_READY_buildy03.md | 2026-10-18 | current | HQ -> lab-host: buildy03 READY (2026-10-18 08:30 EDT)
  <path:HQ_NEWHOST_READY_buildy03.md>
  ssh -p 65022 [redacted-addr] (32 cpu, 125 GB RAM, 128 GB SSD swap, 987 GB disk). Partition 4card, owner lab-host (tokens 4c1,4card,fourcard,mlpv,accpar,att_v1) -- you have full launch/kill aut
- hq:./HQ_NEWHOST_READY_buildy04.md | 2026-10-18 | current | HQ -> lab-host: buildy04 READY (2026-10-18 08:30 EDT)
  <path:HQ_NEWHOST_READY_buildy04.md>
  ssh -p 65022 [redacted-addr] (32 cpu, 125 GB RAM, 128 GB SSD swap, 987 GB disk). Partition 4card, owner lab-host (tokens 4c1,4card,fourcard,mlpv,accpar,att_v1) -- you have full launch/kill aut
- hq:./HQ_NEWHOST_READY_buildy05.md | 2026-10-18 | current | HQ -> lab-host: buildy05 READY (2026-10-18 08:30 EDT)
  <path:HQ_NEWHOST_READY_buildy05.md>
  ssh -p 65022 [redacted-addr] (32 cpu, 125 GB RAM, 128 GB SSD swap, 987 GB disk). Partition 4card, owner lab-host (tokens 4c1,4card,fourcard,mlpv,accpar,att_v1) -- you have full launch/kill aut
- hq:./HQ_NEWHOST_READY_buildzd00.md | 2026-10-18 | current | HQ -> lab-host: buildzd00 READY (2026-10-18 23:36 UTC)
  <path:HQ_NEWHOST_READY_buildzd00.md>
  ssh -p 65022 [redacted-addr] (Vivado at /shared/tools/xilinxinstall/2025.1, 128 GB swap, 134 G install copied from buildxd00/buildzd00 seed, rc=0). Partition 4card, owner lab-host (tokens 4c1
- hq:./HQ_NEWHOST_READY_buildzd01.md | 2026-10-18 | current | HQ -> lab-host: buildzd01 READY (2026-10-18 23:36 UTC)
  <path:HQ_NEWHOST_READY_buildzd01.md>
  ssh -p 65022 [redacted-addr] (Vivado at /shared/tools/xilinxinstall/2025.1, 128 GB swap, 134 G install copied from buildxd00/buildzd00 seed, rc=0). Partition 4card, owner lab-host (tokens 4c1
- hq:./HQ_NEWHOST_READY_buildzd02.md | 2026-10-18 | current | HQ -> lab-host: buildzd02 READY (2026-10-18 23:36 UTC)
  <path:HQ_NEWHOST_READY_buildzd02.md>
  ssh -p 65022 [redacted-addr] (Vivado at /shared/tools/xilinxinstall/2025.1, 128 GB swap, 134 G install copied from buildxd00/buildzd00 seed, rc=0). Partition 4card, owner lab-host (tokens 4c1
- hq:./HQ_NEWHOST_READY_buildzd03.md | 2026-10-18 | current | HQ -> lab-host: buildzd03 READY (2026-10-18 23:36 UTC)
  <path:HQ_NEWHOST_READY_buildzd03.md>
  ssh -p 65022 [redacted-addr] (Vivado at /shared/tools/xilinxinstall/2025.1, 128 GB swap, 134 G install copied from buildxd00/buildzd03 seed, rc=0). Partition 4card, owner lab-host (tokens 4c1

## lab_result (126)

- lab:2026-09-24/E1_D449/REPORT.md | 2026-09-24 | superseded | E1_D449 — LAB_STATE Test 6, 392-node readback discriminator (2026-09-24 23:10) [0xE4B711C8,0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question:** is the w8 d7 (392-node) stale tree readback a RACE (the words arrive late) or a WEDGE/limit (they never arrive)?
- lab:2026-09-24/E1_DEPTH_SWEEP_HUM/REPORT.md | 2026-09-24 | current | E1_DEPTH_SWEEP_HUM — Test 4H: the Test 4 depth sweep on humaneval (2026-09-24 23:16–23:31) [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question (user):** what is the primed vs unprimed SPEED gap on humaneval, per depth? This subsumes build-host-A R3 (primed humaneval at 3:3 / 3:7) in the pair's 50×32 shape; d7 is not in the sweep, and d8
- lab:2026-09-24/E1_DEPTH_SWEEP_HUM/t4h_table.md | 2026-09-24 | current | t4h_table.md
  <path:t4h_table.md>
  
- lab:2026-09-24/E1_HUMANEVAL_W23/REPORT.md | 2026-09-24 | current | E1_HUMANEVAL_W23 — LAB_STATE Test 5 (2026-09-24 23:04–23:09) [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question:** humaneval accept and tok/s of the pair at w2/w3 d2, unprimed, against the GPU bar. This run becomes the pair's humaneval reference.
- lab:2026-09-24/E1_HUM_VS_GPU_50x32/REPORT.md | 2026-09-24 | current | E1_HUM_VS_GPU_50x32 — pair vs GPU-only on humaneval at the pair's exact shape (2026-09-24 23:31–23:4 [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question (user/HQ):** a like-for-like humaneval bar. Pair (Test 4H) vs GPU-only, same shape: w3, depth 2–15, 50 requests, max-new 32, SGLang 0.5.16.
- lab:2026-09-24/E1_HUM_VS_GPU_50x32/hum_vs_gpu_50x32_table.md | 2026-09-24 | current | hum_vs_gpu_50x32_table.md
  <path:hum_vs_gpu_50x32_table.md>
  
- lab:2026-09-24/E1_R2_PRIMEROWS/REPORT.md | 2026-09-24 | current | E1_R2_PRIMEROWS — build-host-A R2 (6c): prompt-KV dose response at w3 d7 (2026-09-24 23:12–23:16) [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question (build-host-A DEPTH_ACCEPT_RCA §6 R2):** how much of the primed accept gain comes from the last N prompt rows? This sizes F0/F2.
- lab:2026-09-24/E1_R2_PRIMEROWS_D15/REPORT.md | 2026-09-24 | current | E1_R2_PRIMEROWS_D15 — R2 repeated at depth 15 (user: "do depth 15 please") (2026-09-24 23:40–23:45) [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question:** does a partial prime still reach most of the full prime's accept at d15?
- lab:2026-09-24/E1_TIME_BREAKDOWN/REPORT.md | 2026-09-24 | current | E1_TIME_BREAKDOWN — where the time goes on E1+CF (2026-09-24 23:00–23:04) [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question (user):** drafting vs verify vs GPU, communication, priming, each GEMM phase, tree/G8.
- lab:2026-09-25/E1_D449_RACEFIX/REPORT.md | 2026-09-25 | current | E1_D449_RACEFIX — Test 6 re-run with the host RACEFIX (w8 d7, npool 392) (2026-09-25 00:24) [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question:** with the host tree-readback race fix (commit 2a37c48e, `TREE_READBACK_RACEFIX_20260925.md`), does the 392-node tree read back cleanly?
- lab:2026-09-25/E1_D449_RACEFIX/TREE_READBACK_RACEFIX_20260925.md | 2026-09-25 | current | Tree-readback RACEFIX, 2026-09-25 (host only, mock-proven, silicon re-run pending)
  <path:TREE_READBACK_RACEFIX_20260925.md>
  **Answer.** The host reads the tree too early. It trusts a 0x78 "done" bit that still belongs to the previous tree. The fix is on the host: after that "done", poll the whole SoA window until both fres
- lab:2026-09-25/E1_PEROP_XVAL/PAIR_PER_OP_XVAL_20260924.md | 2026-09-24 | current | Per-op cross-validation: GPU drafter vs FPGA pair drafter (2026-09-24) [0xE4B711C8,0xE4B711CF]
  <path:PAIR_PER_OP_XVAL_20260924.md>
  **Question.** The pair's depth-3 accept gain is far below the GPU's (mt_bench w3: GPU 2.64 -> 3.19, pair 1.97 -> 2.08; humaneval
- lab:2026-09-25/E1_PEROP_XVAL/REPORT.md | 2026-09-25 | current | E1_PEROP_XVAL — Test 8, per-op cross-validation on E1+CF, w3 d3 (2026-09-25 00:11–00:12) [0xE4B711E1]
  <path:REPORT.md>
  **Question:** do the pair drafter's per-item checkpoints (R3 = h1, G7g, G7u, G11 = a) match the GPU/numpy model op by op on primed input?
- lab:2026-09-25/E1_R3B/REPORT.md | 2026-09-25 | current | E1_R3B — 6d R3b: primed humaneval at the golden's shape (w4 d15, 80 × 1024) — FAILED (2026-09-25 00: [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question:** pair accept at w4 d15 on humaneval 80×1024, primed and unprimed, against the GPU grid (hum d15 w4 fa4 8.15, on the pair's checkpoint).
- lab:2026-09-25/E1_R3B_RACEFIX/REPORT.md | 2026-09-25 | current | E1_R3B_RACEFIX — 6d R3b with the host RACEFIX: w4 d15 humaneval 80×1024, primed then unprimed (2026- [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question:** pair accept at the GPU golden's shape (w4 d15, humaneval, 80 requests, max-new 1024), primed and unprimed, against the GPU grid on the pair's checkpoint (hum d15 w4 fa4 8.15).
- lab:2026-09-25/E1_W8D7_SCORES/REPORT.md | 2026-09-25 | current | E1_W8D7_SCORES — card per-node scores for the w8 d7 golden tree (build-host-A request) (2026-09-25 00:41) [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Question (build-host-A):** separate near-ties from the child RoPE +1 as the cause of the L4–L6 survivor-overlap FAIL (E1_D449_RACEFIX). This needs the card's per-node scores, not just the survivors.
- lab:2026-09-25/E3_FLASH/REPORT.md | 2026-09-25 | current | E3_FLASH — HEAD E3 0xE4B711E3 (E1 + ATTN_PIPE) + TAIL CF 0xE4B711CF, first silicon (2026-09-25 02:17 [0xE4B711CF,0xE4B711E3]
  <path:REPORT.md>
  **Question.** Is E3 output-identical to E1, and does ATTN_PIPE cut attention (G5) time and step time?
- lab:2026-09-25/KVI_EMU/REPORT.md | 2026-09-24 | current | KVI_EMU: can the HEAD use the GPU drafter's own K/V? (2026-09-25)
  <path:REPORT.md>
  (Written by the K/V emulator agent; the harness blocked it from writing this file, so lab-host HQ saved its text verbatim.)
- lab:2026-09-25/KV_DETAIL/REPORT.md | 2026-09-25 | current | KV_DETAIL — LAB_STATE 6e (build-host-A + user), 2026-09-24 23:45 → 09-25 00:08 [0xE4B711CF,0xE4B711E1]
  <path:REPORT.md>
  **Images:** HEAD E1 0xe4b711e1, TAIL CF 0xe4b711cf. **Common regime:** w3, mt_bench, 50 requests, max-new 32, SGLang 0.5.16, fa4, K_C 64,
- lab:2026-09-26/A5F4_FOLD/REPORT.md | 2026-09-26 | current | A5F4 — untested features on HEAD A5 0xE4B711A5 (C7 + G4 embedding fold + SEAM_WIDE 2 burst) + TAIL F [0xE4B711A5]
  <path:REPORT.md>
  - Table verified before upload: 10/10 md5 vs `<path:TABLE_MD5.txt`.>
- lab:2026-09-26/A5F4_PROTOCOL/REPORT.md | 2026-09-26 | current | A5F4_PROTOCOL — standard pair protocol v1 on HEAD A5 0xE4B711A5 (C7 + G4 fold + SEAM_WIDE 2 burst; f [0xE4B711A5]
  <path:REPORT.md>
  37 cells in `A5F4_cells.csv` (fixed columns; concatenates with E7F3 / A8F4 cells). Settings per protocol v1 (50×32, fa4, K_C 64, RACEFIX, KVI off, fastglue off).
- lab:2026-09-26/A8F4_FLASH/REPORT.md | 2026-09-26 | current | A8F4_FLASH — HEAD A8 0xE4B711A8 (E1 + KVI v2 + CHILD_ROPE_FIX + LANES_V 256) + TAIL F4 0xE4B711F4 (F [0xE4B711A8,0xE4B711F4]
  <path:REPORT.md>
  **Verdict so far (KVI OFF): correct (texts 50/50 = E1+CF), and the fastest pair yet on first arms: mt w2 d2 192.51, w3 d2 189.01 @ 1.98.**
- lab:2026-09-26/A8F4_KVI/REPORT.md | 2026-09-26 | current | A8F4_KVI — KV import (KVI v2, wide loader) on silicon, HEAD A8 0xE4B711A8 + TAIL F4 0xE4B711F4, 2026 [0xE4B711A8,0xE4B711F4]
  <path:REPORT.md>
  **Verdict so far: KVI WORKS ON SILICON.** Imported GPU-drafter K/V gives the same draft tree as card priming, 9.4× faster to prime.
- lab:2026-09-26/A8F4_KVI_GRID/REPORT.md | 2026-09-26 | current | A8F4_KVI_GRID — KVI + FASTGLUE=4 (v4.2) sweet-point grid, A8+F4, mt 50×32, SGLv 0.5.16, 2026-09-26 1
  <path:REPORT.md>
  **Verdict: the sweet point is w2 d2 (230.66 tok/s); narrower and shallower is faster at every depth.** Every cell KVI PROOF ok
- lab:2026-09-26/A8F4_KVI_LEVERS/REPORT.md | 2026-09-26 | current | A8F4_KVI_LEVERS — backend A/B with KVI ON (A8+F4, mt 50×32, 0.5.16, fa4 draft default), 2026-09-26 1
  <path:REPORT.md>
  **Verdict: keep the default (T0 = target flashinfer + draft fa4).** Target fa3 is slower; draft fa3 is a tie. Texts 50/50 vs T0 for every
- lab:2026-09-26/ADAPTIVE_TREE/REPORT.md | 2026-09-26 | current | ADAPTIVE_TREE v3 — ran on A5+F4 (HEAD 0xE4B711A5 + TAIL 0xE4B711F4), KVI + FASTGLUE=4 v4.2, mt 50×32 [0xE4B711A5,0xE4B711F4]
  <path:REPORT.md>
  (Planned for A8+F4; run on A5+F4 per HQ fallback since the A5 AC came first. Offline gates ALL GREEN 13/13.)
- lab:2026-09-26/C7F4_FLASH/REPORT.md | 2026-09-26 | current | C7F4_FLASH — HEAD C7 0xE4B711C7 (E3 + KVI v2 + CHILD_ROPE_FIX, 128 lanes) + TAIL F4 0xE4B711F4 (F3 + [0xE4B711C7,0xE4B711F4]
  <path:REPORT.md>
  **Verdict so far (KVI OFF): correct, texts identical to E1+CF, accept up slightly (RoPE fix), TAIL TREE −33%/step exactly as predicted.
- lab:2026-09-26/E7F3_FLASH/REPORT.md | 2026-09-26 | current | E7F3_FLASH — HEAD E7 0xE4B711E7 (E1 + LANES_V 256) + TAIL F3 0xE4B711F3 (CF + T1 G12_OUTST3 + G1 lmd [0xE4B711E7,0xE4B711F3]
  <path:REPORT.md>
  (folder dated 09-26 = the test day; the flashes were 2026-09-25 13:41 / 14:59; pre-patch host backups are in ../../2026-09-25/E7F3_FLASH/prepatch_backup/)
- lab:2026-09-26/GPU_EXTEND_IN_PAIR/REPORT.md | 2026-09-26 | current | GPU_EXTEND_IN_PAIR — the SGLang drafter extend inside the pair step (build-host-A LAB_REQUEST 2026-09-25), [0xE4B711E7,0xE4B711F3]
  <path:REPORT.md>
  **Question.** How much GPU time does the SGLang drafter extend (which KVI would import from) cost per verify step in PAIR mode, and is it
- lab:2026-09-26/HOST_GLUE_V3/REPORT.md | 2026-09-26 | current | HOST_GLUE_V3 on A8+F4 (KVI off) — EAGLE4_PAIR_FASTGLUE=3 (card process + pipelined RPC), 2026-09-26 
  <path:REPORT.md>
  - **Texts (mt w2 d2, 0.5.14):** A (width_ab_20260926_155044), **F3 206.18 @ 1.83** (width_ab_20260926_155229), F2 (width_ab_20260926_155326).
- lab:2026-09-26/HOST_GLUE_V43/REPORT.md | 2026-09-26 | current | HOST_GLUE_V43 — KVI + FASTGLUE=4 v4.3 (real counters), A8+F4, mt d2 50×32, 0.5.16, 2026-09-26 18:03–
  <path:REPORT.md>
  **K4 wins 3/3 rounds at both widths vs K; texts 50/50 every round; KVI PROOF ok.** Confirms v4 (HOST_GLUE_V4_FIXED: K4 234.62 / 222.10).
- lab:2026-09-26/HOST_GLUE_V4_FIXED/REPORT.md | 2026-09-26 | current | HOST_GLUE_V4_FIXED — KVI + EAGLE4_PAIR_FASTGLUE=4 on A8+F4, 2026-09-26 17:05–17:31 (after the lab-host r
  <path:REPORT.md>
  **Verdict: K4 (KVI + FASTGLUE=4) WINS 3/3 rounds at both widths; texts 50/50 vs K every round. NEW PAIR BEST 234.62 tok/s (mt w2 d2).**
- lab:2026-09-26/KVI_PORT/REPORT.md | 2026-09-26 | current | KVI_PORT — KV-import host path for C7 / A8 (host side only; no card touched), 2026-09-26 [0xE4B711A6,0xE4B711A8,0xE4B711AA,0xE4B711C7]
  <path:REPORT.md>
  (Saved by lab-host HQ from the port agent's final report; the harness blocked the agent from writing this file. Logs, diffs, backups and
- lab:2026-09-27/A9BF4_FLASH/REPORT.md | 2026-09-27 | current | A9BF4 — HEAD A9b 0xE4B711A9 (A8 + I12 ARRAY_FAST_TP, 256 lanes, KVI v2 + RoPE fix) + TAIL F4 0xE4B71 [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  - [lanes] build-id 0xe4b711a9 -> GEMV LANES=256; TAIL F4 LANES=128.
- lab:2026-09-27/ACCEPT_CURVE/REPORT.md | 2026-09-27 | current | ACCEPT_CURVE — accept vs depth, pair vs GPU at matched width/depth, 2026-09-27 (user request 17:0x)
  <path:REPORT.md>
  GPU: GPU-only (exec-mode -1, fa3 drafter, WARM + PCG), 50x32, single runs, 17:21-17:33 (gpu/, gpu_accept_curve_*.out).
- lab:2026-09-27/ACCEPT_CURVE_D15/REPORT.md | 2026-09-28 | current | ACCEPT_CURVE_D15 — accept vs depth to d15, GPU half (pair half pending the A5+FA cold AC), 2026-09-2
  <path:REPORT.md>
  Spec: <path:NOTE_HQ_ACCEPT_CURVE_D15.md.> GPU-only: exec-mode -1, fa3 drafter, WARM + PCG, 50x32, 1 run per cell, no cards.
- lab:2026-09-27/ADAPTIVE_FULLSTACK/REPORT.md | 2026-09-27 | current | ADAPTIVE_FULLSTACK — per-step adaptive depth on today's full stack, A9b+F4, 2026-09-27 11:53-12:08
  <path:REPORT.md>
  **Premise correction:** ADAPTIVE_TREE v3 has no live adaptive arm. It taps every tree's card scores (read-only, EAGLE4_PAIR_SCORE_TRACE=1) and fits a
- lab:2026-09-27/CENSUS_A7FB2/REPORT.md | 2026-09-28 | current | Census A7 0xE4B711A7 + FB2 0xE4B711F0, mt w2 d3, 2026-09-27 22:05-22:08 (lab-host) [0xE4B711A7,0xE4B711F0]
  <path:REPORT.md>
  Env = FB2 v1.2 kit stack (KVI FG4 EXTKV GN WARM PCG + FB2 driver) + EAGER + EARLY_A, HOSTPUSH=0. Raw: census_20260927_220542.out.
- lab:2026-09-27/FA_A5FA/REPORT.md | 2026-09-27 | current | FA A/B on A5 0xE4B711A5 + FA 0xE4B711FA — first silicon, 2026-09-27 16:17-16:23 (STOPPED at the EA1  [0xE4B711A5,0xE4B711FA]
  <path:REPORT.md>
  Bring-up (cold AC 16:17): ids 0xe4b711a5 + 0xe4b711fa, qdma roundtrip PASS x2, LANES 128/128, weight/scale drops 0, FRONTIER GATE PASS, VERDICT PASS,
- lab:2026-09-27/FA_HOST_UNIFIED_V11_A5FA/REPORT.md | 2026-09-27 | current | FA A/B with FA_HOST_UNIFIED v1.1 on A5 0xE4B711A5 + FA 0xE4B711FA — the A5 baseline for A7, 2026-09- [0xE4B711A5,0xE4B711FA]
  <path:REPORT.md>
  Full stack in both arms: KVI + FG4 v4.3 + HOSTPUSH v1.1-fa (HPQ0) + EXTKV + GN + WARM + PCG + EAGER. EA0 = EAGER; EA1 = EAGER + EARLY_A.
- lab:2026-09-27/FB2_TS10_A7FB2/REPORT.md | 2026-09-27 | current | TS10 (TREE_STOP lag1 theta -1.0, no SPEC0) on A7 + FB2 — STOPPED on a host crash, 2026-09-27 22:08-2
  <path:REPORT.md>
  Kit FB2_HOST v1.2 lab_fb2_ab.sh ARMS="BASE TS10" CELLS=2:3 HUMCELLS=1:4; dup phase OK; kit md5 rc 0.
- lab:2026-09-27/GLUE2/REPORT.md | 2026-09-27 | current | GLUE2 (host cuts PACK / NOSYNC / PRUNE on the new defaults) on A9b 0xE4B711A9 + F4, 2026-09-27 06:41 [0xE4B711A9]
  <path:REPORT.md>
  Kit md5 14/14. All arms: KVI + FG4 v4.3 + HOSTPUSH HPQ0 + EXTKV. Arms A (defaults), GC (GLUE2 + CHECK every use), G (GLUE2). mt 50x32.
- lab:2026-09-27/GLUE2_V21/REPORT.md | 2026-09-27 | current | GLUE2 v2.1 on A9b 0xE4B711A9 + F4, 2026-09-27 07:02-07:27 (kit <path:v2_1>, md5  [0xE4B711A9]
  <path:REPORT.md>
  v2.1 = graph pre-capture + SELFCHECK off + ROOT on by default. All arms: KVI + FG4 v4.3 + HOSTPUSH HPQ0 + EXTKV (+ LITE probe, so A/B is fair).
- lab:2026-09-27/GPU_PATH/REPORT.md | 2026-09-27 | current | GPU_PATH (EXTKV KV-only GPU drafter extend + CUDA-event timeline + prefill probe) on A9b 0xE4B711A9  [0xE4B711A9]
  <path:REPORT.md>
  Stack in both arms: KVI + FASTGLUE=4 v4.3 + HOSTPUSH v1.1 HPQ0. Kit md5 10/10; offline gates 18/18. Log lab_gpu_path_20260927_044454.out; GPU_PATH_SUMMARY.txt.
- lab:2026-09-27/HOST_PUSH/REPORT.md | 2026-09-27 | current | HOST_PUSH v1 on A9b 0xE4B711A9 + F4 0xE4B711F4, 2026-09-27 03:00-03:20 (kit md5 OK; KVI + FASTGLUE=4 [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  (HQ projection was +0.6-0.9% d2, +3.0-3.5% d7.)
- lab:2026-09-27/HOST_PUSH_V11/REPORT.md | 2026-09-27 | current | HOST_PUSH v1.1 on A9b 0xE4B711A9 + F4 0xE4B711F4, 2026-09-27 03:43-04:10 (kit md5 OK; KVI + FASTGLUE [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  Log: lab_hostpush_ab_20260927_034314.out (this dir). Gates 51/51. Texts 50/50 in all 20 comparisons (texts phase + 3 rounds x 3 cells x 2 HP arms);
- lab:2026-09-27/HOST_PUSH_V12/REPORT.md | 2026-09-27 | current | HOST_PUSH v1.2 on A9b 0xE4B711A9 + F4, 2026-09-27 06:03-06:55 (kit md5 27/27; SLICE host = FE v2; ne [0xE4B711A9]
  <path:REPORT.md>
  Arms: A (no HOSTPUSH), HPQ0 (v1.2, qos off), SK (SKIP_LAST_A only), HSK (HPQ0 + SKIP_LAST_A). mt 50x32, w2 d2/d4/d7, 3 rotated rounds, texts cell w2 d4.
- lab:2026-09-27/HPQ0_FILL/REPORT.md | 2026-09-27 | current | HPQ0 fill — A9b 0xE4B711A9 + F4, KVI + FASTGLUE=4 v4.3 + HOSTPUSH v1.1 HPQ0, mt 50x32, SGLv 0.5.16,  [0xE4B711A9]
  <path:REPORT.md>
  3 rounds, order rotated ([1:2 1:3 2:2 2:3 2:4] / reversed / [2:3 1:2 2:4 2:2 1:3]). All 15 arms: KVI PROOF ok, `[hostpush] armed ... qos=off` 15/15,
- lab:2026-09-27/HPQ0_FILL2/REPORT.md | 2026-09-27 | current | HPQ0 fill 2 — A9b + F4, KVI + FASTGLUE=4 v4.3 + HOSTPUSH v1.1 HPQ0, SGLv 0.5.16, 2026-09-27 04:31-04
  <path:REPORT.md>
  3 rotated rounds per bench; all arms KVI PROOF ok, STATE OK after. Logs: fill_*.out here and ../HPQ0_FILL2_HUM/.
- lab:2026-09-27/HUM_FULLSTACK/REPORT.md | 2026-09-27 | current | HUM_FULLSTACK — humaneval 50x32 on A9b 0xE4B711A9 + F4, full default stack, 2026-09-27 09:21-09:36 [0xE4B711A9]
  <path:REPORT.md>
  Stack: KVI + FASTGLUE=4 v4.3 + HOSTPUSH HPQ0 + EXTKV + GN (GLUE2 v2.1, ROOT off) + WARM (EAGLE4_FA4_WARM=1) + PCG (--enable-piecewise-cuda-graph).
- lab:2026-09-27/LAB_STATE_snapshot.md | 2026-09-27 | current | LAB_STATE — lab-host silicon lab (single source of truth for whoever runs the cards) [0xE4B711A0,0xE4B711A5,0xE4B711A7,0xE4B711A8...]
  <path:LAB_STATE_snapshot.md>
  Updated 2026-09-24 22:50 EDT by lab-host HQ (session ef0dc5). Read this, then the tail of
- lab:2026-09-27/LAB_STATE_snapshot_1703.md | 2026-09-27 | current | LAB_STATE — lab-host silicon lab (single source of truth for whoever runs the cards) [0xE4B711A0,0xE4B711A5,0xE4B711A7,0xE4B711A8...]
  <path:LAB_STATE_snapshot_1703.md>
  Updated 2026-09-24 22:50 EDT by lab-host HQ (session ef0dc5). Read this, then the tail of
- lab:2026-09-27/LEVEL_TIMELINE/REPORT.md | 2026-09-27 | current | LEVEL_TIMELINE — per-level step decomposition on A9b 0xE4B711A9 + F4, 2026-09-27 05:1x-05:33 [0xE4B711A9]
  <path:REPORT.md>
  GPU_PATH kit `timeline report` phases only; TL_CELLS mt w1 d2/d3/d4 + w2 d2/d3/d4; 1 round, 20 requests (REQ=20 confirmed in the SGLang log).
- lab:2026-09-27/NEWDEF_BASELINE/REPORT.md | 2026-09-27 | current | NEW-DEFAULTS baseline — A9b 0xE4B711A9 + F4 0xE4B711F4, 2026-09-27 05:1x-05:3x (reference for FA / H [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  Stack: KVI + FASTGLUE=4 v4.3 + HOSTPUSH v1.1 HPQ0 + EXTKV (= the GPU_PATH X arm env, verbatim), 50x32, SGLv 0.5.16, 3 rotated rounds.
- lab:2026-09-27/PINGPONG_AB/REPORT.md | 2026-09-27 | current | PINGPONG A/B — 2 request streams on one pair (EAGLE4_PAIR_PINGPONG=1), A9b+F4, full stack, 2026-09-2
  <path:REPORT.md>
  Kit <path:PINGPONG> (md5 24/24); gates ALL GREEN (33/33 R, 33/33 W, front 7/7, chain probes GREEN).
- lab:2026-09-27/PINGPONG_AB_V11/REPORT.md | 2026-09-27 | current | PINGPONG A/B -- 2 streams on ONE pair (EAGLE4_PAIR_PINGPONG=1) vs 1-stream serial, 2026-09-27T11:45: [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  HEAD 0xe4b711a9 TAIL 0xe4b711f4 | stack: BASE_PP + EAGLE4_PAIR_FASTGLUE=4 EAGLE4_PAIR_HOSTPUSH=1 EAGLE4_PAIR_EXTKV=1 EAGLE4_PAIR_GLUE2=1 EAGLE4_G2_NO=ROOT EAGLE4_FA4_WARM=1 + EAGLE4_PAIR_KVI=1 EAGLE4_
- lab:2026-09-27/PINGPONG_AB_V12/REPORT.md | 2026-09-27 | current | PINGPONG A/B -- 2 streams on ONE pair (EAGLE4_PAIR_PINGPONG=1) vs 1-stream serial, 2026-09-27T12:21: [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  HEAD 0xe4b711a9 TAIL 0xe4b711f4 | stack: BASE_PP + EAGLE4_PAIR_FASTGLUE=4 EAGLE4_PAIR_HOSTPUSH=1 EAGLE4_PAIR_EXTKV=1 EAGLE4_PAIR_GLUE2=1 EAGLE4_G2_NO=ROOT EAGLE4_FA4_WARM=1 + EAGLE4_PAIR_KVI=1 EAGLE4_
- lab:2026-09-27/PINGPONG_AB_V12_HUM/REPORT.md | 2026-09-27 | current | PINGPONG A/B -- 2 streams on ONE pair (EAGLE4_PAIR_PINGPONG=1) vs 1-stream serial, 2026-09-27T12:30: [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  HEAD 0xe4b711a9 TAIL 0xe4b711f4 | stack: BASE_PP + EAGLE4_PAIR_FASTGLUE=4 EAGLE4_PAIR_HOSTPUSH=1 EAGLE4_PAIR_EXTKV=1 EAGLE4_PAIR_GLUE2=1 EAGLE4_G2_NO=ROOT EAGLE4_FA4_WARM=1 + EAGLE4_PAIR_KVI=1 EAGLE4_
- lab:2026-09-27/PINGPONG_AB_V13/REPORT.md | 2026-09-27 | current | PINGPONG A/B -- 2 streams on ONE pair (EAGLE4_PAIR_PINGPONG=1) vs 1-stream serial, 2026-09-27T12:53: [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  HEAD 0xe4b711a9 TAIL 0xe4b711f4 | stack: BASE_PP + EAGLE4_PAIR_FASTGLUE=4 EAGLE4_PAIR_HOSTPUSH=1 EAGLE4_PAIR_EXTKV=1 EAGLE4_PAIR_GLUE2=1 EAGLE4_G2_NO=ROOT EAGLE4_FA4_WARM=1 + EAGLE4_PAIR_KVI=1 EAGLE4_
- lab:2026-09-27/PINGPONG_AB_V13_MPS/REPORT.md | 2026-09-27 | current | PINGPONG A/B -- 2 streams on ONE pair (EAGLE4_PAIR_PINGPONG=1) vs 1-stream serial, 2026-09-27T16:48: [0xE4B711A5,0xE4B711FA]
  <path:REPORT.md>
  HEAD 0xe4b711a5 TAIL 0xe4b711fa | stack: BASE_PP + EAGLE4_PAIR_FASTGLUE=4 EAGLE4_PAIR_HOSTPUSH=1 EAGLE4_PAIR_EXTKV=1 EAGLE4_PAIR_GLUE2=1 EAGLE4_G2_NO=ROOT EAGLE4_FA4_WARM=1 + EAGLE4_PAIR_KVI=1 EAGLE4_
- lab:2026-09-27/PREFILL_W1W2/REPORT.md | 2026-09-27 | current | PREFILL_W1W2 — A9b 0xE4B711A9 + F4, new defaults (KVI + FG4 v4.3 + HPQ0 + EXTKV), 2026-09-27 07:00-0 [0xE4B711A9]
  <path:REPORT.md>
  Runner prefill_w1w2.sh; log prefill_w1w2_*.out; KVI PROOF ok 9/9; STATE OK after.
- lab:2026-09-27/SESSION_DIGEST_lab-host_20260927.md | 2026-09-27 | current | SESSION DIGEST — lab-host lab, 2026-09-27 (written 21:59 EDT by lab-host-junior, read-only pass) [0xE4B711A5,0xE4B711A7,0xE4B711A9,0xE4B711B4...]
  <path:SESSION_DIGEST_lab-host_20260927.md>
  Sources: LS = `<path:LAB_STATE.md`> (357 lines; identical to the `$W/LAB_STATE.md` copy, `diff -q` clean); R = `<path:`>;
- lab:2026-09-27/SPEC0_A5FA/REPORT.md | 2026-09-27 | current | SPEC0 A/B on A5 0xE4B711A5 + FA 0xE4B711FA, 2026-09-27 18:01-18:20 (run while the A7 cold AC was pen [0xE4B711A5,0xE4B711FA]
  <path:REPORT.md>
  Arms: A = FA_UNIFIED v1.1 + EAGER=1 + EARLY_A=1, HOSTPUSH=0; S = A + EAGLE4_PAIR_SPEC0=1. Rest: KVI + FG4 v4.3 + EXTKV + GN + WARM + PCG + score tap.
- lab:2026-09-27/STALL44/REPORT.md | 2026-09-27 | current | STALL44 — the 4.4 s "Amazon" stall: FA4 JIT compile confirmed; EAGLE4_FA4_WARM=1 fixes it. A9b+F4 + 
  <path:REPORT.md>
  Kit <path:STALL44> (v1 runtime; v2 gate ALL GREEN on lab-host incl. 3 real: OK lines; v1 and v2 stall44.py / sitecustomize.py byte-identical).
- lab:2026-09-27/STALL44/REPORT_V3_PREFILL.md | 2026-09-27 | current | STALL44 v3 PREFILL — piecewise prefill CUDA graph (PCG) for the target, A9b+F4 + H200, 2026-09-27 08
  <path:REPORT_V3_PREFILL.md>
  Kit <path:v3> (md5 6/6). Gate needed TORCH_CUDA_ARCH_LIST=9.0 (see below), then ALL GREEN incl. 5 real: OK lines.
- lab:2026-09-27/STALL44/REPORT_V41_PREFILL2.md | 2026-09-27 | current | STALL44 v4.1 PREFILL2 — drafter prefill graph (DP) and whole-forward target prefill graph (TF) on B 
  <path:REPORT_V41_PREFILL2.md>
  Kit <path:v4.1> (md5 6/6; gates GREEN with TORCH_CUDA_ARCH_LIST=9.0). GPU-only arms: GPU_EXEC_MODE=-1 (v4.1 default).
- lab:2026-09-27/VERIFY_BACKEND_DEPTH/REPORT.md | 2026-09-27 | current | VERIFY_BACKEND_DEPTH — A9b+F4, full stack (KVI + FG4 v4.3 + HPQ0 + EXTKV + GN + WARM + PCG), 2026-09
  <path:REPORT.md>
  KVI PROOF ok 24/24; texts within each backend 50/50 in all 12 comparisons; STATE OK. Log vbd_*.out.
- lab:2026-09-27/WARM_PCG_REBASELINE/REPORT.md | 2026-09-27 | current | WARM+PCG re-baseline — FINAL comparators before A5+FA, 2026-09-27 08:30-08:4x [0xE4B711A9,0xE4B711F4]
  <path:REPORT.md>
  Default stack: KVI + FASTGLUE=4 v4.3 + HOSTPUSH HPQ0 + EXTKV + GN (GLUE2 v2.1, ROOT off) + WARM (EAGLE4_FA4_WARM=1) + PCG (--enable-piecewise-cuda-graph).
- lab:2026-09-27/WIDTH_SWEEP/REPORT.md | 2026-09-27 | current | WIDTH_SWEEP — accept vs verify/FPGA cost over tree width, A9b 0xE4B711A9 + F4, 2026-09-27 05:48-06:0 [0xE4B711A9]
  <path:REPORT.md>
  New defaults (KVI + FASTGLUE=4 v4.3 + HOSTPUSH HPQ0 + EXTKV) + timeline instruments (hostprof LITE + CUDAEV + GP_PROBE), mt 50x32, 2 rounds rotated,
- lab:2026-09-28/FB2_TS10_V13_A7FB2/REPORT.md | 2026-09-28 | current | TS10 re-run with FB2_HOST v1.3 (padded early stops) on A7 0xE4B711A7 + FB2 0xE4B711F0 — PASS, +3.7.. [0xE4B711A7,0xE4B711F0]
  <path:REPORT.md>
  Kit FB2_HOST v1.3 (KIT_MD5 271d1d05 OK; kit gates incl. ts_buckets green on lab-host), lab_fb2_ab.sh ARMS="BASE TS10" CELLS=2:3 HUMCELLS=1:4, 3 rotated rounds.
- lab:2026-09-28/REPRO_A5FA_SPEC0/REPORT.md | 2026-09-28 | current | REPRO of the batch-1 record (A5+FA+SPEC0, mt w2 d3, 307.67) — FAIL on speed, exact; 2026-09-27 23:19
  <path:REPORT.md>
  Fresh cold AC 23:1x; bring-up PASS (ids a5/fa, qdma PASS both, drops 0, FRONTIER PASS). Runner tmp/repro_a5fa.sh = SPEC0_A5FA runner at w2 d3 only,
- lab:2026-09-28/TIMELINE_PAIRS/REPORT.md | 2026-09-28 | current | TIMELINE_PAIRS — per-step host timeline, A5+FA half (A7+FB2 half pending a re-flash), 2026-09-28 01: [0xE4B711A5]
  <path:REPORT.md>
  User "do it" (00:5x): locate the A7+FB2 off-card gap. Instrument = GPU_PATH kit hostprof LITE + CUDAEV + GP_PROBE (HQ NOTE_HQ_TIMELINE_KIT), added to
- lab:2026-09-29/CP1_BDFSL/REPORT.md | 2026-09-29 | current | HX1 0xE4B711BD + FSL 0xE4B711CC — NEW BEST batch-1: S mt w2 d3 323.3 tok/s (+5.1% vs 307.67), exact; [0xE4B711BD,0xE4B711CC]
  <path:REPORT.md>
  BOOT_CLASS=FAST (HEAD-rootport CE at boot=0, runtime GHES=0) WLOAD=see logs
- lab:2026-09-29/CP1_DBFSL_SLOW/REPORT.md | 2026-09-29 | current | HXF-EVAC 0xE4B711DB + FSL 0xE4B711CC — CONS_HOST v1.8, 2026-09-29 09:52-10:58 (lab-host). SLOW boot: rel [0xE4B711CC,0xE4B711DB]
  <path:REPORT.md>
  BOOT_CLASS=SLOW (HEAD-rootport CE at boot=14)
- lab:2026-09-30/CP1_BDFSL_FAST/REPORT.md | 2026-09-30 | current | HX1 0xE4B711BD + FSL 0xE4B711CC — NEW BATCH-1 RECORD: S mt w2 d3 326.23 tok/s (+6.0% vs 307.67), kit [0xE4B711BD,0xE4B711CC]
  <path:REPORT.md>
  BOOT_CLASS=FAST (HEAD-rootport CE at boot=0, runtime GHES=0) — links Gen4 x8
- lab:2026-09-30/VANILLA_H200/REPORT.md | 2026-09-30 | current | VANILLA_H200 — plain (non-speculative) H200 speed vs EAGLE4, same protocol; 2026-09-30 20:26-20:36 (
  <path:REPORT.md>
  NOTE_HQ_VANILLA_H200 (user: "highlight vanilla H200 speed and our speed up of it").
- lab:2026-10-01/CP1_BDAE/REPORT.md | 2026-10-01 | current | HX1 0xE4B711BD + FGI 0xE4B711AE (d197bac9) — valid best TSP12Q mt w2 d7 332.85 (texts exact); TSP12B [0xE4B711AE,0xE4B711BD]
  <path:REPORT.md>
  BOOT_CLASS=FAST (HEAD-rootport CE at boot=0) — first boot after a gated self-poweroff (00:05) + automatic cold AC (back 00:38)
- lab:2026-10-01/CP1_BDAE_FANFAST/REPORT.md | 2026-10-01 | current | HX1 0xE4B711BD + FGI 0xE4B711AE (d197bac9) — FANFAST NO-GO (-4.5%); RELAY_LEAN small win; NEW VALID  [0xE4B711AE,0xE4B711BD]
  <path:REPORT.md>
  BOOT_CLASS=FAST (HEAD-rootport CE at boot=0, 10:41)
- lab:2026-10-01/CP1_BDT75/REPORT.md | 2026-10-01 | current | HX1 0xE4B711BD + TCLK75 TAIL 0xE4B71199 (df7af3df, FGI @ 75 MHz) — TCLK75 GO: NEW VALID BEST TSP12QL [0xE4B71199,0xE4B711BD]
  <path:REPORT.md>
  BOOT_CLASS=FAST (HEAD-rootport CE at boot=0, 14:54)
- lab:2026-10-01/HEADGAP_TL_HX1FGI/REPORT.md | 2026-10-01 | current | REPORT.md
  <path:REPORT.md>
  BOOT_CLASS=FAST (CE 0). HEADGAP_TL v1.1 on HX1 BD + FGI AE, 2026-10-01 11:24-11:26 (lab-host)
- lab:2026-10-03/P1_BASELINE_62_61/REPORT.md | 2026-10-03 | current | Pair-1 baseline on the 4-card box: HX1 0xE4B711BD (HEAD 62:00.0) + TCLK75 0xE4B71199 (TAIL 61:00.0)  [0xE4B71199,0xE4B711BD]
  <path:REPORT.md>
  BOOT_CLASS=FAST by the old 60:01.1 rule (boot CE 0) -- but see the SMI storm below; links Gen4 x8 (V80) / x16 (H200)
- lab:2026-10-03/P1_CSTATE_62_61/REPORT.md | 2026-10-03 | current | C-state / SMI diagnosis on the 4-card box (pair 1: HX1 62 + TCLK75 61) — 2026-10-03 18:3x-18:40 (box
  <path:REPORT.md>
  Question (HQ RCA): is the 2.4x pair-1 slowdown CPU idle-state wake latency? Test: /dev/cpu_dma_latency held at 0 + governor performance, no BIOS/card change.
- lab:2026-10-05/SIBXE_4CARD_STB2/REPORT.md | 2026-10-05 | current | SIBX-E 4-card single stream — first valid silicon cells (lab-host, 2026-10-05 03:0x–03:33) [0xE4B71197,0xE4B711BD]
  <path:REPORT.md>
  **Answer:** the 4-card stream RUNS CORRECTLY but is ~2x SLOWER than the 2-card pair at 2:7: X 50.00 tok/s vs P 96.95 tok/s, texts 50/50, proof ok. 4:7 not measured (kit stopped, gap 4 below).
- lab:2026-10-05/SIBXE_4CARD_STB2_AC3/REPORT.md | 2026-10-05 | current | SIBX-E 4-card A/B after cold AC #3 (lab-host, 2026-10-05 11:56–13:54) — kit v1.23stb2, P-first order
  <path:REPORT.md>
  **Answer:** NO-GO on this platform. 4:7 X still unmeasured: the kit allows only ONE X run per boot (gap 4 also blocks X after X).
- lab:./2026-10-06_disk_cleanup_inventory.md | 2026-10-06 | current | Disk cleanup inventory (READ-ONLY) — lab-host lab-host, 2026-10-06/07
  <path:2026-10-06_disk_cleanup_inventory.md>
  Nothing was deleted/moved. Sizes via `du -x` (GiB-ish, rounded). "refs" = `command grep -F` of the path (or the build-id/dir name where noted) in NOW.md, CLAUDE.md, INDEX.md, TRACKLIST_20260923.md, re
- lab:./2026-10-06_workspace_map_and_restructure_plan.md | 2026-10-06 | current | Workspace map and restructure plan (lab-host, 2026-10-06)
  <path:2026-10-06_workspace_map_and_restructure_plan.md>
  Read-only survey. Nothing was moved, deleted or changed. Sizes are deliberately not covered (separate inventory job).
- lab:2026-10-07/SMI_FIX_RECORDPAIR_104651/REPORT.md | 2026-10-07 | current | SMI storm FIXED at runtime — record pair recovers 140 -> 316.9 tok/s (93% of 339.12) [M] (lab-host, 2026 [0xE4B71199,0xE4B711BD]
  <path:REPORT.md>
  **Root cause [M]:** firmware-first AER on card 01 (01:00.0/.1, root port 00:03.1). Its Gen4 link raises corrected errors; the BIOS handles each in SMM
- lab:2026-10-07/SMI_STORM_PLATFORM_AUDIT.md | 2026-10-07 | current | SMI storm platform audit (lab-host, 2026-10-07, READ-ONLY)
  <path:SMI_STORM_PLATFORM_AUDIT.md>
  Board WRX80E-SAGE SE WIFI II, BIOS 1702, kernel 6.8.0-101. Inputs reused: HOST_GAP_PROBE_v2_093315/{acpi,pci,ipmi,fch}. Live reads taken this session (setpci reads, perf stat, /proc, /sys, xHCI MMIO r
- lab:2026-10-07/SMI_STORM_RESEARCH_web.md | 2026-10-07 | current | SMI storm research (web) 2026-10-07 — lab-host, WRX80E-SAGE SE WIFI II, BIOS 1702
  <path:SMI_STORM_RESEARCH_web.md>
  Honesty note: web search found NO direct report of ls_smi_rx storms on WRX80/TRX40, no ASUS BIOS notes after 1702 (searches returned nothing usable; check ASUS support page manually), and no public do
- lab:2026-10-07/SMI_STORM_TIMELINE.md | 2026-10-07 | current | SMI storm timeline, lab-host (lab-host), forensics 2026-10-07 (read-only)
  <path:SMI_STORM_TIMELINE.md>
  - Last good boot: boot -8, 2026-10-01 14:53 to 23:07 (2 V80 + H200 at 01:00.0; record 339.12 tok/s). Zero slow-NMI lines, zero perf-throttle lines, "tsc: Fast TSC calibration using PIT".
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sx1003_45u1hehc/kit/LAB_PLAN_CP1.md | 2026-10-03 | current | LAB PLAN CP1: HXF 0xE4B711B4 + TAIL (FA | FB2 | FS) on CONS_HOST v1.2 (= v1.1 + TREE_STOP x EAGER/SP [0xE4B711B4,0xE4B711BD,0xE4B711BF,0xE4B711CC...]
  <path:LAB_PLAN_CP1.md>
  Labels: [M] measured, [D] derived from [M], [P] projected. Evidence paths are lab-host (`<path:...`>) or build-host-A (`/shared/...`).
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sx1003_45u1hehc/kit/NOTE_CONS_HOST_v1.22tb.md | 2026-10-03 | current | CONS_HOST v1.22tb: pair BDFs from env + a V80-only guard (HQ, 2026-10-03) [0xE4B71199,0xE4B711BD]
  <path:NOTE_CONS_HOST_v1.22tb.md>
  **Why:** on the 4-card lab-host the V80s sit at 01/42/61/62:00.0, and the **H200 GPU now sits at the old TAIL slot**.
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sx1003_45u1hehc/kit/NOTE_HQ_SIBXE_ONBIT.md | 2026-10-06 | current | HQ → lab-host: CONS_HOST v1.23stb3 fixes the SIBX-E "on" bit (kit gap 4). Use it instead of v1.23stb2 fo [0xE4B71197]
  <path:NOTE_HQ_SIBXE_ONBIT.md>
  **What the bit is (from the RTL of 0xE4B71197, fork tail_sibxe_20261001):**
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sxonbit_9im97amw/kit_new/LAB_PLAN_CP1.md | 2026-10-03 | current | LAB PLAN CP1: HXF 0xE4B711B4 + TAIL (FA | FB2 | FS) on CONS_HOST v1.2 (= v1.1 + TREE_STOP x EAGER/SP [0xE4B711B4,0xE4B711BD,0xE4B711BF,0xE4B711CC...]
  <path:LAB_PLAN_CP1.md>
  Labels: [M] measured, [D] derived from [M], [P] projected. Evidence paths are lab-host (`<path:...`>) or build-host-A (`/shared/...`).
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sxonbit_9im97amw/kit_new/NOTE_CONS_HOST_v1.22tb.md | 2026-10-03 | current | CONS_HOST v1.22tb: pair BDFs from env + a V80-only guard (HQ, 2026-10-03) [0xE4B71199,0xE4B711BD]
  <path:NOTE_CONS_HOST_v1.22tb.md>
  **Why:** on the 4-card lab-host the V80s sit at 01/42/61/62:00.0, and the **H200 GPU now sits at the old TAIL slot**.
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sxonbit_9im97amw/kit_new/NOTE_HQ_SIBXE_ONBIT.md | 2026-10-06 | current | HQ → lab-host: CONS_HOST v1.23stb3 fixes the SIBX-E "on" bit (kit gap 4). Use it instead of v1.23stb2 fo [0xE4B71197]
  <path:NOTE_HQ_SIBXE_ONBIT.md>
  **What the bit is (from the RTL of 0xE4B71197, fork tail_sibxe_20261001):**
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sxonbit_9im97amw/kit_old/LAB_PLAN_CP1.md | 2026-10-03 | current | LAB PLAN CP1: HXF 0xE4B711B4 + TAIL (FA | FB2 | FS) on CONS_HOST v1.2 (= v1.1 + TREE_STOP x EAGER/SP [0xE4B711B4,0xE4B711BD,0xE4B711BF,0xE4B711CC...]
  <path:LAB_PLAN_CP1.md>
  Labels: [M] measured, [D] derived from [M], [P] projected. Evidence paths are lab-host (`<path:...`>) or build-host-A (`/shared/...`).
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sxonbit_9im97amw/kit_old/NOTE_CONS_HOST_v1.22tb.md | 2026-10-03 | current | CONS_HOST v1.22tb: pair BDFs from env + a V80-only guard (HQ, 2026-10-03) [0xE4B71199,0xE4B711BD]
  <path:NOTE_CONS_HOST_v1.22tb.md>
  **Why:** on the 4-card lab-host the V80s sit at 01/42/61/62:00.0, and the **H200 GPU now sits at the old TAIL slot**.
- lab:2026-10-07/TWOPAIR_AC5/PAIR1_62_61/gates/sxonbit_9im97amw/kit_old/NOTE_HQ_SIBXE_ONBIT.md | 2026-10-06 | current | HQ → lab-host: CONS_HOST v1.23stb3 fixes the SIBX-E "on" bit (kit gap 4). Use it instead of v1.23stb2 fo [0xE4B71197]
  <path:NOTE_HQ_SIBXE_ONBIT.md>
  **What the bit is (from the RTL of 0xE4B71197, fork tail_sibxe_20261001):**
- lab:2026-10-07/TWOPAIR_AC5_R2/REPORT.md | 2026-10-07 | current | Two record pairs after cold AC #5 (lab-host, 2026-10-07 06:19–06:40) — HX1 BD 0xE4B711BD + TCLK75 0xE4B7 [0xE4B71199,0xE4B711BD]
  <path:REPORT.md>
  **Answer:** both pairs work and match each other (~140 tok/s, accept 3.62), but each is at ~41-42% of the 339.12 record for this exact image pair [M].
- lab:2026-10-07/TWOPAIR_TOGETHER/REPORT.md | 2026-10-07 | current | Pair 01->42 solo + two-pair-together (batch 2 / aggregate) with the SMI storm fixed (lab-host, 2026-10-0
  <path:REPORT.md>
  Precondition [M]: runtime CE mask on card 01 + RP 00:03.1 in place (SMI 8,920-9,760 / 5 s vs 136k storm). H200 14 MiB used before the run (demo paused).
- lab:./2026-10-07_vivado_consolidation_plan.md | 2026-10-07 | current | Vivado/Vitis 2025.1 consolidation plan (lab-host) - 2026-10-07
  <path:2026-10-07_vivado_consolidation_plan.md>
  Read-only investigation. Nothing was edited, moved, deleted or killed. KEEP = `/shared/tools/xilinxinstall` (nvme1n1p3, group xilinx, setgid).
- lab:2026-10-08/HOSTLEVERS_AB/REPORT.md | 2026-10-08 | current | HOSTLEVERS A/B — 62 HX1 BD 0xE4B711BD -> 42 TCLK75 0xE4B71199, CONS_HOST v1.23stb4g-hostlevers (KIT_ [0xE4B71199,0xE4B711BD]
  <path:REPORT.md>
  Cell F0686 TSP12QL w2 d7 mt_bench, 3 rotated rounds, arm order per round K/Y/G/J/base rotated by lab_cp1; texts 50/50 vs TSP12QL for EVERY arm EVERY round; proof-fail rounds 0.
- lab:2026-10-08/PAIR2_SG_v10v_62_61/REPORT.md | 2026-10-08 | current | PAIR2 SG v1.0v Phase 1 — step report (lab-host, 2026-10-08) [0xE4B71106,0xE4B71107,0xE4B711BD]
  <path:REPORT.md>
  Kit: <path:SG_HOST_v1.0v_on_v1.22tb2> (KIT_MD5 OK on lab-host). HEAD 62:00.0, TAIL 61:00.0. 41 untouched.
- lab:2026-10-08/RECORD_PAIR_AC8_PROFILE/REPORT.md | 2026-10-08 | current | RECORD_PAIR_AC8_PROFILE — 62 HX1 BD 0xE4B711BD -> 42 TCLK75 0xE4B71199, CONS_HOST v1.23stb4d-r1 + HO [0xE4B71199,0xE4B711BD]
  <path:REPORT.md>
  Valid run = **run3/** (11:38). run1/run2 = lab-host setup errors, no card write: run1 TAIL 0x1CC=2 left by the demo (w2) -> bringup WIDTH_FAIL rc 24 (correct guard);
- lab:2026-10-08/REC_62_42_stb4dr1_AC8/REPORT.md | 2026-10-08 | current | Record pair re-measure after AC#8 — 62 HX1 BD 0xE4B711BD -> 42 TCLK75 0xE4B71199, CONS_HOST v1.23stb [0xE4B71199,0xE4B711BD]
  <path:REPORT.md>
  Cell F0686 = TSP12QL w2 d7 mt_bench, x3 rounds [M]. Kit unpacked <path:CONS_HOST_v1.23stb4d-r1> (outer KIT_MD5 OK, inner OK).
- lab:./2026-10-08_build-host-A_ssd_virtual_ram_vivado.md | 2026-10-08 | current | build-host-A: SSD-backed virtual RAM for Vivado 2025.1 full-chip PDI builds (2026-10-08)
  <path:2026-10-08_build-host-A_ssd_virtual_ram_vivado.md>
  Labels: [M] measured, [D] derived, [P] projected. Nothing was applied anywhere.
- lab:./2026-10-08_lab-host_filesystem_design.md | 2026-10-08 | current | lab-host filesystem design (2026-10-08)
  <path:2026-10-08_lab-host_filesystem_design.md>
  Proposal only. Nothing has been moved or erased. Every destructive step below waits for the user's explicit OK.
- lab:./2026-10-11_HANDOFF_lab-host_to_hq_junior.md | 2026-10-11 | current | HANDOFF lab-host (lab-host) -> hq-junior / hq-junior-2, 2026-10-11 01:40 EDT [0xE4B7111B,0xE4B7111C,0xE4B7111D,0xE4B71199...]
  <path:2026-10-11_HANDOFF_lab-host_to_hq_junior.md>
  Labels: [M] measured/read from an artifact, [D] derived, [P] projected. Every path below was `ls`-checked on the named host at writing time unless marked (not checked).
- lab:2026-10-14/DISK_SURVEY_HANDOFF.md | 2026-10-14 | current | Disk survey hand-off (lab-host-main, 2026-10-14 ~06:0x) — for user-c9 / any infra agent. READ-ONLY; n
  <path:DISK_SURVEY_HANDOFF.md>
  Docs: lab-host contract <path:2026-10-08_lab-host_filesystem_design.md> (+ 2026-10-06_disk_cleanup_inventory.md, <path:README.md>, <path:>, ~/SANCTUARY_REORG_PLAN_20260922.md); a
- lab:2026-10-14/THREE_CARD_V3_AC14/FAULT_MODELS_FOR_XSIM.md | 2026-10-14 | current | Fault models for the xsim S_KVL repro (report-forensics fork, lab-host-main, 2026-10-14) — v2
  <path:FAULT_MODELS_FOR_XSIM.md>
  - PRIMARY: M2 (R data never returns on one port after burst 9) and M3 (AR #10 accepted but never answered).
- lab:2026-10-14/THREE_CARD_V3_AC14/IMPL_REPORT_FORENSICS_SKVL.md | 2026-10-14 | current | Card A S_KVL stall — implementation report forensics (lab-host-main fork, 2026-10-14)
  <path:IMPL_REPORT_FORENSICS_SKVL.md>
  Question: do the Vivado synth/impl artifacts explain the silent KVI fan-reader stall on card A
- lab:2026-10-14/THREE_CARD_V3_AC14/KVIFIX_AUDIT.md | 2026-10-14 | current | KVIFIX audit (read-only auditor fork, lab-host-main session, 2026-10-14) [0xE4B7111D]
  <path:KVIFIX_AUDIT.md>
  Target: implementer fork `verilog/forks/head_s3LA_kvifix_20261014/` (card A KVI fan-reader fix: watchdog + abort + per-port debug CSRs + K/V decoupling + CDC xdc fix).
- lab:2026-10-14/THREE_CARD_V3_AC14/RTL_KVI_STALL_ANALYSIS.md | 2026-10-14 | current | Card A S_KVL hang: KVI_FAN reader stall analysis (lab-host, read-only RTL work, 2026-10-14 ~01:05 EDT)
  <path:RTL_KVI_STALL_ANALYSIS.md>
  **Answer.** The card-A RTL has no deadlock that matches the hang. The fan reader, the kernel-side sequencer and the
- lab:2026-10-14/THREE_CARD_V3_AC14/XSIM_SKVL_REPRO.md | 2026-10-14 | current | Card A S_KVL hang — xsim reproduction (lab-host-main xsim fork, 2026-10-14)
  <path:XSIM_SKVL_REPRO.md>
  **Question (user):** use the Vivado build reports and reproduce, in xsim, what we are doing on silicon: the
- lab:2026-10-14/TUNIT_1258/REPORT.md | 2026-10-14 | current | 0xE4B71258 TUNIT on silicon (card 42:00.0, XFL1XVDSKD44) — PASS (2026-10-14 18:17 EDT, lab-host-main)
  <path:REPORT.md>
  Boot after AC#15 (user power-on ~18:16). AMC READY x3 (42/61/62). Image stored on 42: hbms_TUNIT_L0lramT32_0xE4B71258_ec1c83b8.pdi (FLASH_EXIT=0 05:34:51).
- lab:2026-10-14/TUNIT_1258/RUN_ON_SILICON.md | 2026-10-14 | current | 0xE4B71258 TUNIT (Track T constant-fold unit) — silicon check on lab-host card 42 (prepared 2026-10-14 b
  <path:RUN_ON_SILICON.md>
  Image: <path:hbms_TUNIT_L0lramT32_0xE4B71258_ec1c83b8.pdi> (md5 ec1c83b8). Card 42:00.0 = serial XFL1XVDSKD44. NEVER 41:00.0.
- lab:2026-10-15/A2A_0x125D/kit/RUN_ON_lab-host.md | 2026-10-15 | current | RUN_ON_lab-host -- 4-card QSFP ALL-TO-ALL link test, image a2a_link 0xE4B7125D (four_card_a2a, 2026-10-1
  <path:RUN_ON_lab-host.md>
  LINK-ONLY image (no drafter, no HBM engine): 3x Aurora 64B66B GTM x4 25.78125 Gb/s per card on cages 0, 3, 2 (= P1, P4, P3);
- lab:2026-10-15/FOUR_CARD_BRINGUP/NOTES.md | 2026-10-15 | current | FOUR_CARD_BRINGUP 2026-10-15 (lab-host-main) [0xE4B7111B,0xE4B7111C,0xE4B7111D,0xE4B711BD]
  <path:NOTES.md>
  - QDMA card 01 (XFL1UPOZSA25, 0x100 0xe4b711bd HX1 BD): `03_setup_qdma.sh 4 01:00.0` under card_run (OWNER lab-host-main-qdma01) 07:2x -> "Roundtrip test: PASS", /dev/qdma01001-MM-0..3, dma-ctl max QP 14.
- lab:2026-10-15/THREE_CARD_V3_AC16/RCA_KVI0_FIRST_TREE_STALL.md | 2026-10-15 | current | RCA: AC16 3S first-tree stall with KVI=0 ("D2 card loop stalled: tree 0 ... 0x43") [0xE4B7111B,0xE4B7111C,0xE4B7111D]
  <path:RCA_KVI0_FIRST_TREE_STALL.md>
  2026-10-15, lab-host RCA agent. Read-only on the cards: nothing was run on silicon and no BAR was touched. Program: 3-card Llama-3.2-3B
- lab:2026-10-15/TOKEN16_HUNT/REPORT.md | 2026-10-15 | current | TOKEN16_HUNT — 16-bit token/index fields vs the 128,256 vocab (lab-host, 2026-10-15)
  <path:REPORT.md>
  Trigger: HQ research INSIGHTS_FOR_lab-host_20261015.md (TACCEPT false rejects of ids >= 65536 from a 16-bit field). Read-only; nothing edited/built/run on cards.
- lab:2026-10-16/A2A_0x125D/REPORT.md | 2026-10-16 | current | A2A4_LINK 0xE4B7125D on 4 cards — QSFP port discovery (lab-host, AC18 boot 2026-10-16 03:53)
  <path:REPORT.md>
  Cards 01/42/61/62 all UUID 5bbbb166, build 0xe4b7125d, 150 MHz, port info 0xa2a12c03; SMI 2,266/5 s. Raw: a2a.jsonl, *.out here.
- lab:2026-10-16/THREE_CARD_V3_AC17/REPORT.md | 2026-10-16 | current | THREE_CARD_V3 AC17 — first end-to-end 3-card run (kvi0b), lab-host 2026-10-16 [0xE4B7111B]
  <path:REPORT.md>
  Cards: A 42:00.0 0xE4B7121B (reg 0xe4b7111b, PDI 9766f14d) / B 61:00.0 0xE4B7121C / C 62:00.0 0xE4B7121D. Kit THREE_CARD_V3_HOST_v1.4b3_kvi0b_lab-host (KIT_MD5 f0c39a02; gates ALL GREEN 31 runs). KVI_ALL0
- lab:./lab-host_SESSIONS_FOR_REBOOT_20261013.md | 2026-10-13 | current | lab-host sessions and how to reconnect after the cold restart (recorded 2026-10-13 ~12:4x EDT, before po
  <path:lab-host_SESSIONS_FOR_REBOOT_20261013.md>
  - Vivado on lab-host: none running (ps). Mesh lab-host: jobs=0 (mesh_status.sh). GPU 14 MiB, card lock free since 15:01:55 (lab-host-reflash61 rc=0).
- lab:HISTORY/README.md | 2026-09-29 | current | EAGLE4 run history (lab-host) [0xE4B711A7,0xE4B711A9,0xE4B711F0,0xE4B711FA]
  <path:README.md>
  Regenerate: `python3 <path:build_history.py`> (stdlib only, read-only on sources).
- lab:./TO_HQ.md | 2026-10-23 | current | TO_HQ — lab-host lab -> HQ (append-only; newest at the bottom) [0xE4B71106,0xE4B71107,0xE4B71112,0xE4B7111B...]
  <path:TO_HQ.md>
  - File route adopted (hq-junior 21:5x). INDEX.csv backfilled with 15 rows for today's 09-27 results (it had stopped at 09-26 12:16).

## image_bundle (455)

- image:0xE4B71137 |  | current | Image 0xE4B71137 [0xE4B71137]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71138 |  | current | Image 0xE4B71138 [0xE4B71138]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71139 |  | current | Image 0xE4B71139 [0xE4B71139]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B7113B |  | current | Image 0xE4B7113B [0xE4B7113B]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B7113C |  | current | Image 0xE4B7113C [0xE4B7113C]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B7113F |  | current | Image 0xE4B7113F [0xE4B7113F]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71142 |  | current | Image 0xE4B71142 [0xE4B71142]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71152 |  | current | Image 0xE4B71152 [0xE4B71152]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71170 |  | current | Image 0xE4B71170 [0xE4B71170]
  <path:IMAGE.md>
  - image name: **HX3**  build id 0xE4B71170  role HEAD  state: ROUTE FAILED build-vm 2026-10-03 (148,707 congested nets, 182,810 overlaps; no PDI) [M, QM-16]  pdi md5 n/a
- image:0xE4B71171 |  | current | Image 0xE4B71171 [0xE4B71171]
  <path:IMAGE.md>
  - image name: **HX3s**  build id 0xE4B71171  role HEAD  state: QUEUED 2026-10-02 (mesh waiter)  pdi md5 n/a
- image:0xE4B71173 |  | current | Image 0xE4B71173 [0xE4B71173]
  <path:IMAGE.md>
  - image name: **HX3T**  build id 0xE4B71173  role HEAD  state: PDI written 2026-10-04 08:38 (build dir); FAIL timing usr_2  pdi md5 fd93c295
- image:0xE4B71174 |  | current | Image 0xE4B71174 [0xE4B71174]
  <path:IMAGE.md>
  - image name: **HX3T0**  build id 0xE4B71174  role HEAD  state: QUEUED 2026-10-03 (mesh waiter); stage /shared/build/hx3_stage  pdi md5 n/a
- image:0xE4B71175 |  | current | Image 0xE4B71175 [0xE4B71175]
  <path:IMAGE.md>
  - image name: **HX2QR**  build id 0xE4B71175  role HEAD  state: LANDED 2026-10-04 18:40 build-vm; FAIL timing - not relayed  pdi md5 2b91bf25
- image:0xE4B71175_HX2QRdc |  | current | Image 0xE4B71175_HX2QRdc [0xE4B71175_HX2QRdc]
  <path:IMAGE.md>
  - image name: **HX2QRdc**  build id 0xE4B71175  role HEAD  state: FAIL QM-36 2026-10-10 12:27 EDT: 101x Route 35-557 pin-assignment overlap, post-route WNS -9.628 [M]; STEP5 no FPT PDI, mesh rc=1  pdi
- image:0xE4B71175_HX2QRpb |  | current | Image 0xE4B71175_HX2QRpb [0xE4B71175_HX2QRpb]
  <path:IMAGE.md>
  - image name: **HX2QRpb**  build id 0xE4B71175  role HEAD  state: FAILED 2026-10-06 21:25 EDT build-vm (wall 61,860 s, peak 90.8 GB): route_design Route 35-2 4756 node overlaps + 100x Route 35-557; NO 
- image:0xE4B71175_HX2QRs2 |  | current | Image 0xE4B71175_HX2QRs2 [0xE4B71175_HX2QRs2]
  <path:IMAGE.md>
  - image name: **HX2QRs2**  build id 0xE4B71175  role HEAD  state: LANDED 2026-10-06 00:35 EDT build-vm (wall 80,002 s, peak 96.5 GB); timing FAIL (not relayed)  pdi md5 9faf26d1ae9d091bd62d69227e211e50
- image:0xE4B71176 |  | current | Image 0xE4B71176 [0xE4B71176]
  <path:IMAGE.md>
  - image name: **HX2QK**  build id 0xE4B71176  role HEAD  state: LANDED 2026-10-04 14:18 build-vm; FAIL timing - not relayed (HX2QK watcher verdict; mirrored by QM-20)  pdi md5 d678ea74
- image:0xE4B7117C |  | current | Image 0xE4B7117C [0xE4B7117C]
  <path:IMAGE.md>
  - image name: **HXFP**  build id 0xE4B7117C  role HEAD  state: FAIL-quarantined  pdi md5 2ad3d5e8
- image:0xE4B7117D |  | current | Image 0xE4B7117D [0xE4B7117D]
  <path:IMAGE.md>
  - image name: **HXFP75**  build id 0xE4B7117D  role HEAD  state: withdrawn  pdi md5 n/a
- image:0xE4B7117F |  | current | Image 0xE4B7117F [0xE4B7117F]
  <path:IMAGE.md>
  - image name: **HDC75**  build id 0xE4B7117F  role HEAD  state: LANDED 2026-10-02 12:57; FAIL (info only, never relay)  pdi md5 5b6a5cc8
- image:0xE4B7117F_HDC75b |  | current | Image 0xE4B7117F_HDC75b [0xE4B7117F_HDC75b]
  <path:IMAGE.md>
  - image name: **HDC75b**  build id 0xE4B7117F  role HEAD  state: LANDED 2026-10-03 07:52; PASS  pdi md5 ca7768d3
- image:0xE4B71180 |  | current | Image 0xE4B71180 [0xE4B71180]
  <path:IMAGE.md>
  - image name: **HX3_x**  build id 0xE4B71180  role HEAD  state: RELEASED 2026-10-02 02:48 before admission (id collides with historic 0x1180/0x1181 uses in scripts/); moved to 0x70/0x71  pdi md5 n/a
- image:0xE4B71181 |  | current | Image 0xE4B71181 [0xE4B71181]
  <path:IMAGE.md>
  - image name: **HX3s_x**  build id 0xE4B71181  role HEAD  state: RELEASED 2026-10-02 02:48 before admission (id collides with historic 0x1180/0x1181 uses in scripts/); moved to 0x70/0x71  pdi md5 n/a
- image:0xE4B71183 |  | current | Image 0xE4B71183 [0xE4B71183]
  <path:IMAGE.md>
  - image name: **HDC80**  build id 0xE4B71183  role HEAD  state: ROUTED FAIL [M]: WNS -3.822 / TNS -68,709 (inter-clock -3.8 = v1 derive bug: ACDC XDC not imported; intra usr_0 -2.855 real @80; usr_1 -
- image:0xE4B71184 |  | current | Image 0xE4B71184 [0xE4B71184]
  <path:IMAGE.md>
  - image name: **HX2Q**  build id 0xE4B71184  role HEAD  state: LANDED 2026-10-03 18:20 EDT build-vm; FAIL timing (not relayed)  pdi md5 370a195d
- image:0xE4B71185 |  | current | Image 0xE4B71185 [0xE4B71185]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71186 |  | current | Image 0xE4B71186 [0xE4B71186]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71187 |  | current | Image 0xE4B71187 [0xE4B71187]
  <path:IMAGE.md>
  - image name: **HQ_BH**  build id 0xE4B71187  role HEAD  state: reserved  pdi md5 n/a
- image:0xE4B71188 |  | current | Image 0xE4B71188 [0xE4B71188]
  <path:IMAGE.md>
  - image name: **TQ_BH**  build id 0xE4B71188  role TAIL  state: reserved  pdi md5 n/a
- image:0xE4B7118D |  | current | Image 0xE4B7118D [0xE4B7118D]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71190 |  | current | Image 0xE4B71190 [0xE4B71190]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71192 |  | current | Image 0xE4B71192 [0xE4B71192]
  <path:IMAGE.md>
  - role: unknown (no PDI in <path:pdi_inbox>)
- image:0xE4B71194 |  | current | Image 0xE4B71194 [0xE4B71194]
  <path:IMAGE.md>
  - image name: **G5GPP**  build id 0xE4B71194  role HEAD  state: LANDED 2026-10-01 14:51Z build-vm; FAIL timing (not relayed)  pdi md5 d5da6abb
- image:0xE4B71194_G5GPPax |  | current | Image 0xE4B71194_G5GPPax [0xE4B71194_G5GPPax]
  <path:IMAGE.md>
  - image name: **G5GPPax**  build id 0xE4B71194  role HEAD  state: LANDED 2026-10-01 18:42Z build-vm; FAIL (19 gate FAILs incl. OSPREYCHECK + build-manifest strap mismatch; not relayed)  pdi md5 3a91398e
- image:0xE4B71194_G5GPPinc |  | current | Image 0xE4B71194_G5GPPinc [0xE4B71194_G5GPPinc]
  <path:IMAGE.md>
  - image name: **G5GPPinc**  build id 0xE4B71194  role HEAD  state: LANDED 2026-10-02 14:38; FAIL timing (not relayed)  pdi md5 da9714c5
- image:0xE4B71194_G5GPPmd |  | current | Image 0xE4B71194_G5GPPmd [0xE4B71194_G5GPPmd]
  <path:IMAGE.md>
  - image name: **G5GPPmd**  build id 0xE4B71194  role HEAD  state: LANDED 2026-10-01 08:34Z; FAIL timing (not relayed)  pdi md5 88407bf1
- image:0xE4B71194_G5GPPnd |  | current | Image 0xE4B71194_G5GPPnd [0xE4B71194_G5GPPnd]
  <path:IMAGE.md>
  - image name: **G5GPPnd**  build id 0xE4B71194  role HEAD  state: LANDED 2026-10-01 ~13:50Z; QUARANTINED 19 FAIL; FAIL timing  pdi md5 f8ed90ce
- image:0xE4B71194_G5GPPx |  | current | Image 0xE4B71194_G5GPPx [0xE4B71194_G5GPPx]
  <path:IMAGE.md>
  - image name: **G5GPPx**  build id 0xE4B71194  role HEAD  state: LANDED 2026-10-01 19:20Z build-vm; FAIL (19 gate FAILs incl. OSPREYCHECK + BCAST slack -0.174 + build-manifest mismatch; not relayed)  p
- image:0xE4B71196 |  | current | Image 0xE4B71196 [0xE4B71196]
  <path:IMAGE.md>
  - image name: **SIBX**  build id 0xE4B71196  role TAIL  state: LANDED 2026-10-01 14:52Z build-vm; PASS; relayed HOLD_4CARD  pdi md5 ae08eb73
- image:0xE4B71197 |  | current | Image 0xE4B71197 [0xE4B71197]
  <path:IMAGE.md>
  - image name: **SIBXE**  build id 0xE4B71197  role TAIL  state: LANDED 2026-10-01 15:59 EDT build-host-A; PASS -> relayed FOUR_CARD/B_SIBX (4-card only)  pdi md5 638ded66
- image:0xE4B71197_SIBXEb |  | current | Image 0xE4B71197_SIBXEb [0xE4B71197_SIBXEb]
  <path:IMAGE.md>
  - image name: **SIBXEb**  build id 0xE4B71197  role TAIL  state: RELAYED 2026-10-05 03:29 lab-host pdi_inbox/FOUR_CARD/B_SIBX as the named SIBX-E spare (create-only, md5 OK on lab-host; NOTE_QM21_SIBXEb_SP
- image:0xE4B71197_SIBXEs3 |  | current | Image 0xE4B71197_SIBXEs3 [0xE4B71197_SIBXEs3]
  <path:IMAGE.md>
  - image name: **SIBXEs3**  build id 0xE4B71197  role TAIL  state: LANDED 2026-10-05 09:36 EDT build-host-A (mesh rc=0 wall 22,296 s, peak 90.6 GB); PASS -> RELAYED 09:38 lab-host FOUR_CARD/B_SIBX as spare #2
- image:0xE4B71198 |  | current | Image 0xE4B71198 [0xE4B71198]
  <path:IMAGE.md>
  - image name: **TCLK80**  build id 0xE4B71198  role TAIL  state: LANDED 2026-10-01 22:05 EDT build-vm; HOLD -> lab-host pdi_inbox/HOLD_TCLK80 (BCAST margin 0.262<0.3 needs HQ waiver)  pdi md5 f453a376
- image:0xE4B71198_TCLK80px |  | current | Image 0xE4B71198_TCLK80px [0xE4B71198_TCLK80px]
  <path:IMAGE.md>
  - image name: **TCLK80px**  build id 0xE4B71198  role TAIL  state: LANDED 2026-10-06 03:16 EDT build-host-M (wall 26,963 s, peak 63.0 GB); same HOLD class as TCLK80: BCAST slack 0.262 < 0.3 (identical to f
- image:0xE4B71199 |  | current | Image 0xE4B71199 [0xE4B71199]
  <path:IMAGE.md>
  - image name: **TCLK75s2**  build id 0xE4B71199  role TAIL  state: LANDED 2026-10-02 15:58; clean SPARE (not relayed)  pdi md5 906e2bb4
- image:0xE4B71199_TCLK75 |  | current | Image 0xE4B71199_TCLK75 [0xE4B71199_TCLK75]
  <path:IMAGE.md>
  - image name: **TCLK75**  build id 0xE4B71199  role TAIL  state: LANDED 2026-10-01 build-host-M (TCLK75md); PASS; relayed pdi_inbox  pdi md5 df7af3df
- image:0xE4B71199_TCLK75px |  | current | Image 0xE4B71199_TCLK75px [0xE4B71199_TCLK75px]
  <path:IMAGE.md>
  - image name: **TCLK75px**  build id 0xE4B71199  role TAIL  state: LANDED 2026-10-06 06:33 EDT build-vm (wall 51,440 s, peak 62.4 GB); clean SPARE, not relayed (< TCLK75 +0.016)  pdi md5 9a2180608de84da
- image:0xE4B71199_TCLK75rd |  | current | Image 0xE4B71199_TCLK75rd [0xE4B71199_TCLK75rd]
  <path:IMAGE.md>
  - image name: **TCLK75rd**  build id 0xE4B71199  role TAIL  state: LANDED 2026-10-06 10:28 EDT build-host-M (wall 25,681 s, peak 61.7 GB); PASS zero margin; SPARE, not relayed (< TCLK75 +0.016)  pdi md5 a7
- image:0xE4B7119A |  | current | Image 0xE4B7119A [0xE4B7119A]
  <path:IMAGE.md>
  - image name: **HXD0**  build id 0xE4B7119A  role HEAD  state: WITHDRAWN 2026-10-01 11:41 EDT (never admitted) — HX1 census rules out slow DOT_E3 (F0675-F0677)  pdi md5 n/a
- image:0xE4B7119B |  | current | Image 0xE4B7119B [0xE4B7119B]
  <path:IMAGE.md>
  - image name: **HXLP0**  build id 0xE4B7119B  role HEAD  state: LANDED 2026-10-01 23:05 EDT build-vm; PASS -> PENDING relay to lab-host HOLD_HEDGE (lab-host unreachable 23:2x; auto-relay armed)  pdi md5 abf
- image:0xE4B7119C |  | current | Image 0xE4B7119C [0xE4B7119C]
  <path:IMAGE.md>
  - image name: **HCLK70**  build id 0xE4B7119C  role HEAD  state: v1 FAILED (BD 41-758 ap_clk_hbm; step29 order) -> v2 (clk_out3 300) staged hclk_stage_v2, queued by QM-15  pdi md5 n/a
- image:0xE4B7119D |  | current | Image 0xE4B7119D [0xE4B7119D]
  <path:IMAGE.md>
  - image name: **HCLK75**  build id 0xE4B7119D  role HEAD  state: v1 FAILED (BD 41-758 ap_clk_hbm; step29 order) -> v2 (clk_out3 300) staged hclk_stage_v2, queued by QM-15  pdi md5 n/a
- image:0xE4B7119E |  | current | Image 0xE4B7119E [0xE4B7119E]
  <path:IMAGE.md>
  - image name: **HQ_D2**  build id 0xE4B7119E  role HEAD  state: LANDED 2026-10-02 07:06; FAIL timing (not relayed)  pdi md5 645151ce
- image:0xE4B7119E_HQ_D2 var A |  | current | Image 0xE4B7119E_HQ_D2 var A [0xE4B7119E_HQ_D2 var A]
  <path:0xE4B7119E_HQ_D2> var A/IMAGE.md
  - image name: **HQ_D2 var A**  build id 0xE4B7119E  role HEAD  state: EARLY-LANDED 2026-10-03 from build dir (EVAC_TOOTH still running); PASS timing  pdi md5 9c262304
- image:0xE4B7119E_HQ_D2 var B |  | current | Image 0xE4B7119E_HQ_D2 var B [0xE4B7119E_HQ_D2 var B]
  <path:0xE4B7119E_HQ_D2> var B/IMAGE.md
  - image name: **HQ_D2 var B**  build id 0xE4B7119E  role HEAD  state: EARLY-LANDED 2026-10-03 from build dir (EVAC_TOOTH still running); PASS timing  pdi md5 3c004cd6
- image:0xE4B7119E_HQ_D2 var C |  | current | Image 0xE4B7119E_HQ_D2 var C [0xE4B7119E_HQ_D2 var C]
  <path:0xE4B7119E_HQ_D2> var C/IMAGE.md
  - image name: **HQ_D2 var C**  build id 0xE4B7119E  role HEAD  state: LANDED 2026-10-03 17:54; PASS timing; link/REF census pending  pdi md5 3e830fe7
- image:0xE4B7119F |  | current | Image 0xE4B7119F [0xE4B7119F]
  <path:IMAGE.md>
  - image name: **TQ_D2**  build id 0xE4B7119F  role TAIL  state: LANDED 2026-10-02 02:16; PASS (gates+census) but NOT relayed: pair HEAD HQ_D2 FAIL  pdi md5 e759f0a6
- image:0xE4B7119F_TQ_D2s2 |  | current | Image 0xE4B7119F_TQ_D2s2 [0xE4B7119F_TQ_D2s2]
  <path:IMAGE.md>
  - image name: **TQ_D2s2**  build id 0xE4B7119F  role TAIL  state: RELAYED 2026-10-05 16:52 EDT create-only FOUR_CARD/D2 + NOTE_QM22_TQ_D2s2_SPARE_fbfad46f.md (kit D2_HOST v1.0), md5sum -c OK on lab-host; 
- image:0xE4B711A0 |  | current | Image 0xE4B711A0 [0xE4B711A0]
  <path:IMAGE.md>
  - image name: **AG**  build id 0xE4B711A0  role HEAD  state: BUILT - HOLD (timing FAIL); AG-r physopt rescue (build-vm)  pdi md5 d623fde8e75034a50429b7f9a0da0455
- image:0xE4B711A1 |  | current | Image 0xE4B711A1 [0xE4B711A1]
  <path:IMAGE.md>
  - image name: **A1**  build id 0xE4B711A1  role HEAD  state: FAILED  pdi md5 n/a
- image:0xE4B711A2 |  | current | Image 0xE4B711A2 [0xE4B711A2]
  <path:IMAGE.md>
  - image name: **A2**  build id 0xE4B711A2  role HEAD  state: built  pdi md5 3f7e1ec1
- image:0xE4B711A3 |  | current | Image 0xE4B711A3 [0xE4B711A3]
  <path:IMAGE.md>
  - image name: **A3**  build id 0xE4B711A3  role HEAD  state: built  pdi md5 1fd8740a
- image:0xE4B711A4 |  | current | Image 0xE4B711A4 [0xE4B711A4]
  <path:IMAGE.md>
  - image name: **KS1**  build id 0xE4B711A4  role HEAD  state: UN-QUARANTINED 2026-09-28 (r2 gate GREEN); re-queue from existing handoff_ks1/run_0xE4B711A4_ps.sh (QM, low priority)  pdi md5 n/a
- image:0xE4B711A5 |  | current | Image 0xE4B711A5 [0xE4B711A5]
  <path:IMAGE.md>
  - image name: **A5**  build id 0xE4B711A5  role HEAD  state: tested  pdi md5 1f934249
- image:0xE4B711A6 |  | current | Image 0xE4B711A6 [0xE4B711A6]
  <path:IMAGE.md>
  - image name: **A6**  build id 0xE4B711A6  role HEAD  state: FAILED  pdi md5 n/a
- image:0xE4B711A7 |  | current | Image 0xE4B711A7 [0xE4B711A7]
  <path:IMAGE.md>
  - image name: **A7**  build id 0xE4B711A7  role HEAD  state: LANDED -> lab-host  pdi md5 b31e381c0211e65487dce3c5363a9a86
- image:0xE4B711A8 |  | current | Image 0xE4B711A8 [0xE4B711A8]
  <path:IMAGE.md>
  - image name: **A8**  build id 0xE4B711A8  role HEAD  state: tested  pdi md5 97e717ed
- image:0xE4B711A8_A8b |  | current | Image 0xE4B711A8_A8b [0xE4B711A8_A8b]
  <path:IMAGE.md>
  - image name: **A8b**  build id 0xE4B711A8  role HEAD  state: BUILT - NOT relayed  pdi md5 898b77a42d1cc2524424883ceacf2609
- image:0xE4B711A9 |  | current | Image 0xE4B711A9 [0xE4B711A9]
  <path:IMAGE.md>
  - image name: **A9**  build id 0xE4B711A9  role HEAD  state: LANDED -> lab-host (twin of A9b)  pdi md5 2f082a9c2c69dd9e1248c1b7aff9f6ef
- image:0xE4B711A9_A9b |  | current | Image 0xE4B711A9_A9b [0xE4B711A9_A9b]
  <path:IMAGE.md>
  - image name: **A9b**  build id 0xE4B711A9  role HEAD  state: tested  pdi md5 939c6942
- image:0xE4B711AA |  | current | Image 0xE4B711AA [0xE4B711AA]
  <path:IMAGE.md>
  - image name: **AA**  build id 0xE4B711AA  role HEAD  state: FAILED  pdi md5 n/a
- image:0xE4B711AA_AAb |  | current | Image 0xE4B711AA_AAb [0xE4B711AA_AAb]
  <path:IMAGE.md>
  - image name: **AAb**  build id 0xE4B711AA  role HEAD  state: FAILED  pdi md5 n/a
- image:0xE4B711AB |  | current | Image 0xE4B711AB [0xE4B711AB]
  <path:IMAGE.md>
  - image name: **AB**  build id 0xE4B711AB  role HEAD  state: building  pdi md5 n/a
- image:0xE4B711AC |  | current | Image 0xE4B711AC [0xE4B711AC]
  <path:IMAGE.md>
  - image name: **AC**  build id 0xE4B711AC  role HEAD  state: FAILED route (not requeued)  pdi md5 n/a
- image:0xE4B711AD |  | current | Image 0xE4B711AD [0xE4B711AD]
  <path:IMAGE.md>
  - image name: **AD**  build id 0xE4B711AD  role HEAD  state: withdrawn (coverage plan)  pdi md5 n/a
- image:0xE4B711AE |  | current | Image 0xE4B711AE [0xE4B711AE]
  <path:IMAGE.md>
  - image name: **FGIpx**  build id 0xE4B711AE  role TAIL  state: RELAYED 2026-10-01 03:30Z (lab-host md5+magic verified) = PREFERRED FGI  pdi md5 746d2f94
- image:0xE4B711AE_FGI |  | current | Image 0xE4B711AE_FGI [0xE4B711AE_FGI]
  <path:IMAGE.md>
  - image name: **FGI**  build id 0xE4B711AE  role TAIL  state: RELAYED 2026-10-01 02:24Z (lab-host verified); SUPERSEDED by FGIpx 746d2f94 (spare)  pdi md5 d197bac9
- image:0xE4B711AE_FGIb |  | current | Image 0xE4B711AE_FGIb [0xE4B711AE_FGIb]
  <path:IMAGE.md>
  - image name: **FGIb**  build id 0xE4B711AE  role TAIL  state: LANDED 2026-10-01 ~04:45Z (spare; NOT relayed)  pdi md5 e66e9db4
- image:0xE4B711AF |  | current | Image 0xE4B711AF [0xE4B711AF]
  <path:IMAGE.md>
  - image name: **AF**  build id 0xE4B711AF  role HEAD  state: LANDED build-vm 11:19Z - TIMING FAIL (not relayed)  pdi md5 09ed7a98
- image:0xE4B711B0 |  | current | Image 0xE4B711B0 [0xE4B711B0]
  <path:IMAGE.md>
  - image name: **B0**  build id 0xE4B711B0  role TAIL  state: superseded  pdi md5 fc156ddb
- image:0xE4B711B1 |  | current | Image 0xE4B711B1 [0xE4B711B1]
  <path:IMAGE.md>
  - image name: **HS3**  build id 0xE4B711B1  role HEAD  state: r7 build-host-M copy ROUTE FAIL 2026-09-29 23:02Z [M]: Route 35-2, 195,491 node overlaps (AG spread recipe on), no PDI; build-vm copy ROUTE FAIL
- image:0xE4B711B2 |  | current | Image 0xE4B711B2 [0xE4B711B2]
  <path:IMAGE.md>
  - image name: **QH1**  build id 0xE4B711B2  role HEAD  state: queued FIRST (head of host queue)  pdi md5 n/a
- image:0xE4B711B3 |  | current | Image 0xE4B711B3 [0xE4B711B3]
  <path:IMAGE.md>
  - image name: **QH1b**  build id 0xE4B711B3  role HEAD  state: WITHDRAWN pre-admission 2026-09-28 (superseded by CP1Q; specs stale_state/withdrawn_packall_*.json)  pdi md5 n/a
- image:0xE4B711B4 |  | current | Image 0xE4B711B4 [0xE4B711B4]
  <path:IMAGE.md>
  - image name: **HXF**  build id 0xE4B711B4  role HEAD  state: LANDED build-vm 16:32Z - TIMING FAIL (not relayed)  pdi md5 901566f2
- image:0xE4B711B6 |  | current | Image 0xE4B711B6 [0xE4B711B6]
  <path:IMAGE.md>
  - image name: **B6**  build id 0xE4B711B6  role HEAD  state: LANDED -> lab-host  pdi md5 13b72e9021765d82de0691985dd727a9
- image:0xE4B711B7 |  | current | Image 0xE4B711B7 [0xE4B711B7]
  <path:IMAGE.md>
  - image name: **HX2b**  build id 0xE4B711B7  role HEAD  state: LANDED build-host-A 18:09Z - TIMING FAIL (not relayed) - DEAD  pdi md5 84ec7432
- image:0xE4B711B8 |  | current | Image 0xE4B711B8 [0xE4B711B8]
  <path:IMAGE.md>
  - image name: **HM300**  build id 0xE4B711B8  role HEAD  state: DROPPED (not started; HQ 2026-09-28: HXF supersedes; handoff kept as reference)  pdi md5 n/a
- image:0xE4B711B8_HXFQ |  | current | Image 0xE4B711B8_HXFQ [0xE4B711B8_HXFQ]
  <path:IMAGE.md>
  - image name: **HXFQ**  build id 0xE4B711B8  role HEAD  state: queued build-vm behind HXF2 (requeued 06:04Z, *_dg40.sh; first admission 05:11Z failed DISK GUARD 46G<60G)  pdi md5 n/a
- image:0xE4B711B9 |  | current | Image 0xE4B711B9 [0xE4B711B9]
  <path:IMAGE.md>
  - image name: **HMF300**  build id 0xE4B711B9  role HEAD  state: SUPERSEDED by HXF 0xE4B711B4 (not admitted; handoff kept as reference)  pdi md5 n/a
- image:0xE4B711BA |  | current | Image 0xE4B711BA [0xE4B711BA]
  <path:IMAGE.md>
  - image name: **AG2**  build id 0xE4B711BA  role HEAD  state: FAILED route  pdi md5 n/a
- image:0xE4B711BB |  | current | Image 0xE4B711BB [0xE4B711BB]
  <path:IMAGE.md>
  - image name: **HM**  build id 0xE4B711BB  role HEAD  state: BUILT - HOLD (timing FAIL) -> lab-host pdi_inbox/HOLD_HM (DO NOT FLASH); HM-r2 re-route FAILED to close (WNS -0.038, 268 fail, no PDI, 06:25
- image:0xE4B711BC |  | current | Image 0xE4B711BC [0xE4B711BC]
  <path:IMAGE.md>
  - image name: **HMF**  build id 0xE4B711BC  role HEAD  state: designing  pdi md5 n/a
- image:0xE4B711BD |  | current | Image 0xE4B711BD [0xE4B711BD]
  <path:IMAGE.md>
  - image name: **HX1**  build id 0xE4B711BD  role HEAD  state: LANDED -> lab-host pdi_inbox (QM relay 2026-09-28T09:35:11Z, md5+magic verified on lab-host) - FLASH CANDIDATE (user OK; pair FB2)  pdi md5 ce7
- image:0xE4B711BD_HX1ax |  | current | Image 0xE4B711BD_HX1ax [0xE4B711BD_HX1ax]
  <path:IMAGE.md>
  - image name: **HX1ax**  build id 0xE4B711BD  role HEAD  state: LANDED 2026-09-30 15:31Z (spare; NOT relayed)  pdi md5 aa4d4464
- image:0xE4B711BD_HX1nr |  | current | Image 0xE4B711BD_HX1nr [0xE4B711BD_HX1nr]
  <path:IMAGE.md>
  - image name: **HX1nr**  build id 0xE4B711BD  role HEAD  state: LANDED 2026-10-06 04:08 EDT build-host-A (wall 26,089 s, peak 72.9 GB); clean SPARE, not relayed (< BD +0.004)  pdi md5 ccb0fec5caeede17d2feb6
- image:0xE4B711BD_HX1pd |  | current | Image 0xE4B711BD_HX1pd [0xE4B711BD_HX1pd]
  <path:IMAGE.md>
  - image name: **HX1pd**  build id 0xE4B711BD  role HEAD  state: LANDED 2026-09-30 ~23:44Z (spare; NOT relayed)  pdi md5 9445d202
- image:0xE4B711BD_HX1px |  | current | Image 0xE4B711BD_HX1px [0xE4B711BD_HX1px]
  <path:IMAGE.md>
  - image name: **HX1px**  build id 0xE4B711BD  role HEAD  state: LANDED 2026-10-01 ~04:05Z (spare; NOT relayed)  pdi md5 eb86dffe
- image:0xE4B711BE |  | current | Image 0xE4B711BE [0xE4B711BE]
  <path:IMAGE.md>
  - image name: **HX2**  build id 0xE4B711BE  role HEAD  state: DROPPED 2026-09-28 (mesh job stale-killed by senior after 16 h heavy_impl flock wait, no Vivado ever ran; superseded by HXF2 0xE4B711C0 = 
- image:0xE4B711BF |  | current | Image 0xE4B711BF [0xE4B711BF]
  <path:IMAGE.md>
  - image name: **HS2**  build id 0xE4B711BF  role HEAD  state: RELAYED lab-host pdi_inbox 12:46Z (auto v3)  pdi md5 6625b706
- image:0xE4B711C0 |  | current | Image 0xE4B711C0 [0xE4B711C0]
  <path:IMAGE.md>
  - image name: **HXF2**  build id 0xE4B711C0  role HEAD  state: LANDED build-vm 00:46Z - TIMING FAIL -0.106 (CLOSED, no respin)  pdi md5 f5ae3e8d
- image:0xE4B711C1 |  | current | Image 0xE4B711C1 [0xE4B711C1]
  <path:IMAGE.md>
  - image name: **C1**  build id 0xE4B711C1  role HEAD  state: superseded (untested)  pdi md5 4800fce0
- image:0xE4B711C2 |  | current | Image 0xE4B711C2 [0xE4B711C2]
  <path:IMAGE.md>
  - image name: **HXS**  build id 0xE4B711C2  role HEAD  state: DROPPED 2026-09-29 (HQ decision)  pdi md5 n/a
- image:0xE4B711C3 |  | current | Image 0xE4B711C3 [0xE4B711C3]
  <path:IMAGE.md>
  - image name: **C3**  build id 0xE4B711C3  role HEAD  state: superseded (untested)  pdi md5 7969e900
- image:0xE4B711C4 |  | current | Image 0xE4B711C4 [0xE4B711C4]
  <path:IMAGE.md>
  - image name: **HXL**  build id 0xE4B711C4  role HEAD  state: RESERVED 2026-09-28 by slr1_ff build arm; gates in flight  pdi md5 n/a
- image:0xE4B711C5 |  | current | Image 0xE4B711C5 [0xE4B711C5]
  <path:IMAGE.md>
  - image name: **C5**  build id 0xE4B711C5  role HEAD  state: built  pdi md5 56dcfd62
- image:0xE4B711C6 |  | current | Image 0xE4B711C6 [0xE4B711C6]
  <path:IMAGE.md>
  - image name: **HXE**  build id 0xE4B711C6  role HEAD  state: RESERVED 2026-09-28 by slr1_ff EVAC arm; STATIC GREEN; queue request in QM_LOG  pdi md5 n/a
- image:0xE4B711C7 |  | current | Image 0xE4B711C7 [0xE4B711C7]
  <path:IMAGE.md>
  - image name: **C7**  build id 0xE4B711C7  role HEAD  state: tested  pdi md5 4e9014f6
- image:0xE4B711C8 |  | current | Image 0xE4B711C8 [0xE4B711C8]
  <path:IMAGE.md>
  - image name: **C8**  build id 0xE4B711C8  role HEAD  state: tested  pdi md5 f52a300e
- image:0xE4B711C9 |  | current | Image 0xE4B711C9 [0xE4B711C9]
  <path:IMAGE.md>
  - image name: **C9**  build id 0xE4B711C9  role HEAD  state: superseded  pdi md5 n/a
- image:0xE4B711CA |  | current | Image 0xE4B711CA [0xE4B711CA]
  <path:IMAGE.md>
  - image name: **CA**  build id 0xE4B711CA  role TAIL  state: superseded  pdi md5 1f3c06d3
- image:0xE4B711CB |  | current | Image 0xE4B711CB [0xE4B711CB]
  <path:IMAGE.md>
  - image name: **CB**  build id 0xE4B711CB  role TAIL  state: superseded (untested)  pdi md5 03a728f4
- image:0xE4B711CC |  | current | Image 0xE4B711CC [0xE4B711CC]
  <path:IMAGE.md>
  - image name: **FSL**  build id 0xE4B711CC  role TAIL  state: RELAYED lab-host pdi_inbox 23:34Z (auto v3)  pdi md5 bdc29613
- image:0xE4B711CC_FSLax |  | current | Image 0xE4B711CC_FSLax [0xE4B711CC_FSLax]
  <path:IMAGE.md>
  - image name: **FSLax**  build id 0xE4B711CC  role TAIL  state: LANDED 2026-09-30 21:22Z; auto-relayed 21:24Z (lab-host md5 verified) but SPARE -- NOTE_QM10 says preferred FSL stays bdc29613  pdi md5 e034
- image:0xE4B711CD |  | current | Image 0xE4B711CD [0xE4B711CD]
  <path:IMAGE.md>
  - image name: **CD**  build id 0xE4B711CD  role HEAD  state: superseded (untested)  pdi md5 e3e10459
- image:0xE4B711CE |  | current | Image 0xE4B711CE [0xE4B711CE]
  <path:IMAGE.md>
  - image name: **FSLQ**  build id 0xE4B711CE  role TAIL  state: handoff ready 2026-09-28; waits FSL gates (fl0/fl1m/flrm in flight); TAIL of CP1Q v2 meanwhile = FBRQ 0xE4B711EB  pdi md5 n/a
- image:0xE4B711CF |  | current | Image 0xE4B711CF [0xE4B711CF]
  <path:IMAGE.md>
  - image name: **CF**  build id 0xE4B711CF  role TAIL  state: on card  pdi md5 3e3bc79a
- image:0xE4B711D0 |  | current | Image 0xE4B711D0 [0xE4B711D0]
  <path:IMAGE.md>
  - image name: **D0**  build id 0xE4B711D0  role probe  state: superseded (untested)  pdi md5 n/a
- image:0xE4B711D1 |  | current | Image 0xE4B711D1 [0xE4B711D1]
  <path:IMAGE.md>
  - image name: **D1**  build id 0xE4B711D1  role probe  state: superseded (untested)  pdi md5 n/a
- image:0xE4B711D2 |  | current | Image 0xE4B711D2 [0xE4B711D2]
  <path:IMAGE.md>
  - image name: **D2**  build id 0xE4B711D2  role probe  state: superseded (untested)  pdi md5 dac1864a
- image:0xE4B711D3 |  | current | Image 0xE4B711D3 [0xE4B711D3]
  <path:IMAGE.md>
  - image name: **D3**  build id 0xE4B711D3  role probe  state: superseded (untested)  pdi md5 e9688708
- image:0xE4B711D4 |  | current | Image 0xE4B711D4 [0xE4B711D4]
  <path:IMAGE.md>
  - image name: **HQ_LV2**  build id 0xE4B711D4  role HEAD  state: LANDED 2026-10-01 20:37 EDT build-vm; RELEASED 2026-10-01 ~22:00 EDT -> lab-host FOUR_CARD/C_QSFP/LV2 (second QSFP option; host CP1Q_LINK
- image:0xE4B711D4_HQ_LV2s2 |  | current | Image 0xE4B711D4_HQ_LV2s2 [0xE4B711D4_HQ_LV2s2]
  <path:IMAGE.md>
  - image name: **HQ_LV2s2**  build id 0xE4B711D4  role HEAD  state: LANDED 2026-10-03 11:32; clean spare  pdi md5 db431757
- image:0xE4B711D4_HQ_LV2s3 |  | current | Image 0xE4B711D4_HQ_LV2s3 [0xE4B711D4_HQ_LV2s3]
  <path:IMAGE.md>
  - image name: **HQ_LV2s3**  build id 0xE4B711D4  role HEAD  state: EARLY-LANDED 2026-10-03 21:58 from build dir (PDI 19:04); clean spare  pdi md5 99f03245
- image:0xE4B711D4_HQ_LV2s4 |  | current | Image 0xE4B711D4_HQ_LV2s4 [0xE4B711D4_HQ_LV2s4]
  <path:IMAGE.md>
  - image name: **HQ_LV2s4**  build id 0xE4B711D4  role HEAD  state: EARLY-LANDED 2026-10-04 05:17 from build dir; clean spare  pdi md5 edf49d15
- image:0xE4B711D5 |  | current | Image 0xE4B711D5 [0xE4B711D5]
  <path:IMAGE.md>
  - image name: **TQ_LV2md**  build id 0xE4B711D5  role TAIL  state: LANDED 2026-10-01 20:20 EDT build-host-M; RELEASED 2026-10-01 ~22:00 EDT -> lab-host FOUR_CARD/C_QSFP/LV2 (second QSFP option; host CP1Q_LI
- image:0xE4B711D5_TQ_LV2ae |  | current | Image 0xE4B711D5_TQ_LV2ae [0xE4B711D5_TQ_LV2ae]
  <path:IMAGE.md>
  - image name: **TQ_LV2ae**  build id 0xE4B711D5  role TAIL  state: RELAYED 2026-10-05 03:02 lab-host pdi_inbox/FOUR_CARD/QSFP_MP (relay_when_up.sh RELAY_OK, md5sum -c OK on lab-host)  pdi md5 340a751b
- image:0xE4B711D6 |  | current | Image 0xE4B711D6 [0xE4B711D6]
  <path:IMAGE.md>
  - image name: **D6**  build id 0xE4B711D6  role probe  state: quarantined  pdi md5 69da3cf2
- image:0xE4B711D7 |  | current | Image 0xE4B711D7 [0xE4B711D7]
  <path:IMAGE.md>
  - image name: **D7**  build id 0xE4B711D7  role probe  state: fork-ready  pdi md5 n/a
- image:0xE4B711D8 |  | current | Image 0xE4B711D8 [0xE4B711D8]
  <path:IMAGE.md>
  - image name: **FG8I1Q**  build id 0xE4B711D8  role TAIL  state: v2 LANDED build-host-A 23:27Z md5 d126462c (XSA guard worked: full STEP5) but LINK TIMING FAIL [M]: WNS -21.766 / WHS -10.016 (clk_pl_0->clk_
- image:0xE4B711D9 |  | current | Image 0xE4B711D9 [0xE4B711D9]
  <path:IMAGE.md>
  - image name: **D9**  build id 0xE4B711D9  role probe  state: D9 FAILED place (GT pblock); D9b queued build-vm lowest priority (design D parked)  pdi md5 n/a
- image:0xE4B711DA |  | current | Image 0xE4B711DA [0xE4B711DA]
  <path:IMAGE.md>
  - image name: **DA**  build id 0xE4B711DA  role probe  state: PARKED (design D, HQ 2026-09-29) - DA3 no PDI  pdi md5 n/a
- image:0xE4B711DB |  | current | Image 0xE4B711DB [0xE4B711DB]
  <path:IMAGE.md>
  - image name: **HXF-EVAC**  build id 0xE4B711DB  role HEAD  state: RELAYED lab-host pdi_inbox 04:58Z (QM manual) - CP1 HEAD, NEXT flash after HX1+FSL  pdi md5 6bbd3675
- image:0xE4B711DB_DBax |  | current | Image 0xE4B711DB_DBax [0xE4B711DB_DBax]
  <path:IMAGE.md>
  - image name: **DBax**  build id 0xE4B711DB  role HEAD  state: RELAYED 2026-09-30 (lab-host verified) = PREFERRED DB  pdi md5 11b23a5b
- image:0xE4B711DB_DBex |  | current | Image 0xE4B711DB_DBex [0xE4B711DB_DBex]
  <path:IMAGE.md>
  - image name: **DBex**  build id 0xE4B711DB  role HEAD  state: LANDED 2026-09-30 08:54Z (spare; NOT relayed)  pdi md5 024e6fe7
- image:0xE4B711DC |  | current | Image 0xE4B711DC [0xE4B711DC]
  <path:IMAGE.md>
  - image name: **HXEQ v3**  build id 0xE4B711DC  role HEAD  state: LANDED 2026-09-30 12:35Z - LINK CDC TIMING FAIL (not relayed)  pdi md5 7ca6edce
- image:0xE4B711DC_HQ_DC_v4 |  | current | Image 0xE4B711DC_HQ_DC_v4 [0xE4B711DC_HQ_DC_v4]
  <path:IMAGE.md>
  - image name: **HQ_DC_v4**  build id 0xE4B711DC  role HEAD  state: LANDED 2026-10-01 08:26Z; RELEASED 2026-10-01 ~16:57 EDT -> FOUR_CARD/C_QSFP (link gates G3/G4/G6/G6b PASS; G1/G2 waived = TQ_EB_v4; 
- image:0xE4B711DC_HQ_DC_v4s2 |  | current | Image 0xE4B711DC_HQ_DC_v4s2 [0xE4B711DC_HQ_DC_v4s2]
  <path:IMAGE.md>
  - image name: **HQ_DC_v4s2**  build id 0xE4B711DC  role HEAD  state: LANDED 2026-10-02 05:55; clean SPARE (not relayed)  pdi md5 6036cfe0
- image:0xE4B711DC_HXEQ |  | current | Image 0xE4B711DC_HXEQ [0xE4B711DC_HXEQ]
  <path:IMAGE.md>
  - image name: **HXEQ**  build id 0xE4B711DC  role HEAD  state: v2 ENDED 2026-09-29 ~15:00Z NO PDI (link XSA wall: amc.elf missing; xdc v1 = LINK CDC FAIL); v3 (xdc v2.1 + XSA guard, SLOT gCP1Q_dc_v3) 
- image:0xE4B711DD |  | current | Image 0xE4B711DD [0xE4B711DD]
  <path:IMAGE.md>
  - image name: **HXI1**  build id 0xE4B711DD  role HEAD  state: RELAYED 2026-09-29 (lab-host verified)  pdi md5 24971883
- image:0xE4B711DD_HXI1ax |  | current | Image 0xE4B711DD_HXI1ax [0xE4B711DD_HXI1ax]
  <path:IMAGE.md>
  - image name: **HXI1ax**  build id 0xE4B711DD  role HEAD  state: LANDED 2026-09-30 09:54Z (NOT relayed; ties DD)  pdi md5 621bab66
- image:0xE4B711DE |  | current | Image 0xE4B711DE [0xE4B711DE]
  <path:IMAGE.md>
  - image name: **HXQI1**  build id 0xE4B711DE  role HEAD  state: HANDOFF READY handoff/hxqi1 (db5c3ff1); HOLD for HXEQ D9 GT/BUFG_GT_SYNC fix; TAIL = FG8I1Q D8  pdi md5 n/a
- image:0xE4B711DF |  | current | Image 0xE4B711DF [0xE4B711DF]
  <path:IMAGE.md>
  - image name: **FG8I1**  build id 0xE4B711DF  role TAIL  state: RELAYED 2026-09-29 (lab-host verified)  pdi md5 f91b1ee2
- image:0xE4B711DF_FG8I1ax |  | current | Image 0xE4B711DF_FG8I1ax [0xE4B711DF_FG8I1ax]
  <path:IMAGE.md>
  - image name: **FG8I1ax**  build id 0xE4B711DF  role TAIL  state: LANDED 2026-09-30 05:43Z (spare; NOT relayed)  pdi md5 2dab83a9
- image:0xE4B711E0 |  | current | Image 0xE4B711E0 [0xE4B711E0]
  <path:IMAGE.md>
  - image name: **FSI1**  build id 0xE4B711E0  role TAIL  state: reserved fallback, NOT NEEDED (FG8I1 GREEN)  pdi md5 n/a
- image:0xE4B711E1 |  | current | Image 0xE4B711E1 [0xE4B711E1]
  <path:IMAGE.md>
  - image name: **E1**  build id 0xE4B711E1  role HEAD  state: on card  pdi md5 7b02c367
- image:0xE4B711E2 |  | current | Image 0xE4B711E2 [0xE4B711E2]
  <path:IMAGE.md>
  - image name: **E2**  build id 0xE4B711E2  role HEAD  state: superseded (untested)  pdi md5 f5757eb2
- image:0xE4B711E3 |  | current | Image 0xE4B711E3 [0xE4B711E3]
  <path:IMAGE.md>
  - image name: **E3**  build id 0xE4B711E3  role HEAD  state: on card  pdi md5 809219ea
- image:0xE4B711E4 |  | current | Image 0xE4B711E4 [0xE4B711E4]
  <path:IMAGE.md>
  - image name: **FG8L**  build id 0xE4B711E4  role TAIL  state: RELAYED 2026-09-29 (lab-host verified; build-host-M copy)  pdi md5 7124bb8c
- image:0xE4B711E4_FG8L spare |  | current | Image 0xE4B711E4_FG8L spare [0xE4B711E4_FG8L spare]
  <path:0xE4B711E4_FG8L> spare/IMAGE.md
  - image name: **FG8L spare**  build id 0xE4B711E4  role TAIL  state: LANDED 2026-09-29 16:05Z (spare; NOT relayed)  pdi md5 cf32b81b
- image:0xE4B711E4_FG8Lax |  | current | Image 0xE4B711E4_FG8Lax [0xE4B711E4_FG8Lax]
  <path:IMAGE.md>
  - image name: **FG8Lax**  build id 0xE4B711E4  role TAIL  state: RELAYED 2026-09-30 (lab-host verified) = PREFERRED FG8L  pdi md5 26a5f241
- image:0xE4B711E4_FG8Lpd |  | current | Image 0xE4B711E4_FG8Lpd [0xE4B711E4_FG8Lpd]
  <path:IMAGE.md>
  - image name: **FG8Lpd**  build id 0xE4B711E4  role TAIL  state: LANDED 2026-10-01 ~01:55Z (spare; NOT relayed)  pdi md5 bad9ee9d
- image:0xE4B711E5 |  | current | Image 0xE4B711E5 [0xE4B711E5]
  <path:IMAGE.md>
  - image name: **E5**  build id 0xE4B711E5  role HEAD  state: superseded (untested)  pdi md5 n/a
- image:0xE4B711E6 |  | current | Image 0xE4B711E6 [0xE4B711E6]
  <path:IMAGE.md>
  - image name: **HX1I1**  build id 0xE4B711E6  role HEAD  state: RELAYED 2026-09-29 (lab-host verified)  pdi md5 018f1265
- image:0xE4B711E6_HX1I1 spare |  | current | Image 0xE4B711E6_HX1I1 spare [0xE4B711E6_HX1I1 spare]
  <path:0xE4B711E6_HX1I1> spare/IMAGE.md
  - image name: **HX1I1 spare**  build id 0xE4B711E6  role HEAD  state: LANDED 2026-09-29 (spare; NOT relayed)  pdi md5 1a063b0f
- image:0xE4B711E6_HX1I1ax |  | current | Image 0xE4B711E6_HX1I1ax [0xE4B711E6_HX1I1ax]
  <path:IMAGE.md>
  - image name: **HX1I1ax**  build id 0xE4B711E6  role HEAD  state: RELAYED 2026-09-30 (lab-host verified) = PREFERRED E6  pdi md5 14947b32
- image:0xE4B711E7 |  | current | Image 0xE4B711E7 [0xE4B711E7]
  <path:IMAGE.md>
  - image name: **E7**  build id 0xE4B711E7  role HEAD  state: tested  pdi md5 907f44ee
- image:0xE4B711E8 |  | current | Image 0xE4B711E8 [0xE4B711E8]
  <path:IMAGE.md>
  - image name: **HXPMS**  build id 0xE4B711E8  role HEAD  state: LANDED 2026-10-03 16:14; PASS  pdi md5 33dca36f
- image:0xE4B711E9 |  | current | Image 0xE4B711E9 [0xE4B711E9]
  <path:IMAGE.md>
  - image name: **FG8**  build id 0xE4B711E9  role TAIL  state: id reserved by QM-junior 2026-09-28 for agent aa53eede126dfee32 (tail_fg8); handoff pending /shared/scratch/tail_fg8/handoff/fg8/; queue a
- image:0xE4B711EA |  | current | Image 0xE4B711EA [0xE4B711EA]
  <path:IMAGE.md>
  - image name: **HQ_MP2**  build id 0xE4B711EA  role HEAD  state: LANDED official 2026-10-03 23:15 (= early c6d3fd00); HOLD hold-timing  pdi md5 c6d3fd00
- image:0xE4B711EA_HQ_MP2hm |  | current | Image 0xE4B711EA_HQ_MP2hm [0xE4B711EA_HQ_MP2hm]
  <path:IMAGE.md>
  - image name: **HQ_MP2hm**  build id 0xE4B711EA  role HEAD  state: LANDED 2026-10-05 ~18:14 EDT build-host-M (wall 30,425 s, peak 84.5 GB); PASS; RELAYED 19:48 FOUR_CARD/QSFP_MP as MP2 HEAD SPARE + NOTE_QM
- image:0xE4B711EA_HQ_MP2s2 |  | current | Image 0xE4B711EA_HQ_MP2s2 [0xE4B711EA_HQ_MP2s2]
  <path:IMAGE.md>
  - image name: **HQ_MP2s2**  build id 0xE4B711EA  role HEAD  state: LANDED 2026-10-04 05:07; PASS; kit MP2_HOST v1.0 staged - MP2 HEAD of record (supersedes HQ_MP2 c6d3fd00)  pdi md5 d9d26efc
- image:0xE4B711EA_HQ_MP2s3 |  | current | Image 0xE4B711EA_HQ_MP2s3 [0xE4B711EA_HQ_MP2s3]
  <path:IMAGE.md>
  - image name: **HQ_MP2s3**  build id 0xE4B711EA  role HEAD  state: LANDED 2026-10-05 10:13 EDT build-host-M (PDI 09:47; async evac census still running); FAIL hold - not relayed  pdi md5 b891af00
- image:0xE4B711EB |  | current | Image 0xE4B711EB [0xE4B711EB]
  <path:IMAGE.md>
  - image name: **TQ_EB_v4**  build id 0xE4B711EB  role TAIL  state: LANDED 2026-10-01 ~03:30Z; HELD (1 new class E3-SLR1-fit, awaiting HQ waiver)  pdi md5 492a12fe
- image:0xE4B711EB_FBRQ |  | current | Image 0xE4B711EB_FBRQ [0xE4B711EB_FBRQ]
  <path:IMAGE.md>
  - image name: **FBRQ**  build id 0xE4B711EB  role TAIL  state: v3 ENDED build-vm 01:30Z at the XSA wall; RESCUED (qm6 recover) PDI 8ed371bb [M]; LINK TIMING FAIL: WNS -16.079 / TNS -70,224 (clk_pl_0->c
- image:0xE4B711EB_TQ_EB_v4s2 |  | current | Image 0xE4B711EB_TQ_EB_v4s2 [0xE4B711EB_TQ_EB_v4s2]
  <path:IMAGE.md>
  - image name: **TQ_EB_v4s2**  build id 0xE4B711EB  role TAIL  state: LANDED 2026-10-02 05:25; clean SPARE (not relayed)  pdi md5 e135bd27
- image:0xE4B711EC |  | current | Image 0xE4B711EC [0xE4B711EC]
  <path:IMAGE.md>
  - image name: **FBR**  build id 0xE4B711EC  role TAIL  state: not needed (FB2 landed)  pdi md5 n/a
- image:0xE4B711ED |  | current | Image 0xE4B711ED [0xE4B711ED]
  <path:IMAGE.md>
  - image name: **FAR**  build id 0xE4B711ED  role TAIL  state: not needed (FA LANDED 12f04f1b)  pdi md5 n/a
- image:0xE4B711EE |  | current | Image 0xE4B711EE [0xE4B711EE]
  <path:IMAGE.md>
  - image name: **FN8**  build id 0xE4B711EE  role TAIL  state: RELAYED lab-host pdi_inbox (QM manual) 20:2xZ  pdi md5 d62a4796
- image:0xE4B711EF |  | current | Image 0xE4B711EF [0xE4B711EF]
  <path:IMAGE.md>
  - image name: **FS**  build id 0xE4B711EF  role TAIL  state: RELAYED lab-host pdi_inbox 11:22Z (auto v3)  pdi md5 9ec0f2a2
- image:0xE4B711F0 |  | current | Image 0xE4B711F0 [0xE4B711F0]
  <path:IMAGE.md>
  - image name: **FB2**  build id 0xE4B711F0  role TAIL  state: LANDED -> lab-host  pdi md5 c11dc669d22ed842dc3f49b5176e2447
- image:0xE4B711F1 |  | current | Image 0xE4B711F1 [0xE4B711F1]
  <path:IMAGE.md>
  - image name: **F1**  build id 0xE4B711F1  role TAIL  state: superseded (untested)  pdi md5 036a7261
- image:0xE4B711F2 |  | current | Image 0xE4B711F2 [0xE4B711F2]
  <path:IMAGE.md>
  - image name: **HXR3E**  build id 0xE4B711F2  role HEAD  state: LANDED 2026-10-02 19:19; PASS -> HOLD lab-host pdi_inbox/FLEET/M1 (relay-when-up; flash after farm xsim FM1_R3E)  pdi md5 5c09e0ec
- image:0xE4B711F2_HXR3Es2 |  | current | Image 0xE4B711F2_HXR3Es2 [0xE4B711F2_HXR3Es2]
  <path:IMAGE.md>
  - image name: **HXR3Es2**  build id 0xE4B711F2  role HEAD  state: LANDED 2026-10-04 11:44 build-vm; PASS (QM-19): 0 new/0 dropped FAIL vs HXR3E 5c09e0ec; SPARE (margin +0.001 < HXR3E +0.005); relayed FL
- image:0xE4B711F3 |  | current | Image 0xE4B711F3 [0xE4B711F3]
  <path:IMAGE.md>
  - image name: **F3**  build id 0xE4B711F3  role TAIL  state: on card (tested)  pdi md5 693a35da
- image:0xE4B711F4 |  | current | Image 0xE4B711F4 [0xE4B711F4]
  <path:IMAGE.md>
  - image name: **F4**  build id 0xE4B711F4  role TAIL  state: on card (tested with A5)  pdi md5 85af2587
- image:0xE4B711F5 |  | current | Image 0xE4B711F5 [0xE4B711F5]
  <path:IMAGE.md>
  - image name: **F5**  build id 0xE4B711F5  role TAIL  state: LANDED -> lab-host  pdi md5 3726e648fda7bdcfbf0c156d61a39489
- image:0xE4B711F6 |  | current | Image 0xE4B711F6 [0xE4B711F6]
  <path:IMAGE.md>
  - image name: **F6**  build id 0xE4B711F6  role TAIL  state: LANDED -> lab-host  pdi md5 84230281c0bbb22c444471d1e16e14c2
- image:0xE4B711F7 |  | current | Image 0xE4B711F7 [0xE4B711F7]
  <path:IMAGE.md>
  - image name: **F7**  build id 0xE4B711F7  role TAIL  state: withdrawn (coverage plan)  pdi md5 n/a
- image:0xE4B711F8 |  | current | Image 0xE4B711F8 [0xE4B711F8]
  <path:IMAGE.md>
  - image name: **F8**  build id 0xE4B711F8  role TAIL  state: LANDED -> lab-host pdi_inbox (QM relay 2026-09-28 ~09:20Z, md5+magic verified on lab-host) - NO FLASH UNLESS ASKED  pdi md5 ed6b90c6fb6bb8935513
- image:0xE4B711F9 |  | current | Image 0xE4B711F9 [0xE4B711F9]
  <path:IMAGE.md>
  - image name: **F9**  build id 0xE4B711F9  role TAIL  state: KILLED (build-host-A restart 18:53 EDT, 1h into impl); not re-queued: TREE_STOP/LX shipped in FB2 (landed)  pdi md5 n/a
- image:0xE4B711FA |  | current | Image 0xE4B711FA [0xE4B711FA]
  <path:IMAGE.md>
  - image name: **FA**  build id 0xE4B711FA  role TAIL  state: LANDED -> lab-host pdi_inbox  pdi md5 12f04f1bb423072df3900072b8cabaf0
- image:0xE4B711FB |  | current | Image 0xE4B711FB [0xE4B711FB]
  <path:IMAGE.md>
  - image name: **QT1**  build id 0xE4B711FB  role TAIL  state: WITHDRAWN pre-admission 2026-09-28 (superseded by CP1Q; specs stale_state/withdrawn_packall_*.json)  pdi md5 n/a
- image:0xE4B711FC |  | current | Image 0xE4B711FC [0xE4B711FC]
  <path:IMAGE.md>
  - image name: **QT1b**  build id 0xE4B711FC  role TAIL  state: WITHDRAWN pre-admission 2026-09-28 (superseded by CP1Q; specs stale_state/withdrawn_packall_*.json)  pdi md5 n/a
- image:0xE4B711FD |  | current | Image 0xE4B711FD [0xE4B711FD]
  <path:IMAGE.md>
  - image name: **FD**  build id 0xE4B711FD  role TAIL  state: LANDED -> lab-host pdi_inbox (auto-relay 05:13Z, md5+magic verified on lab-host)  pdi md5 930e64e6e2d059a4e18e7abb48456bcd
- image:0xE4B711FE |  | current | Image 0xE4B711FE [0xE4B711FE]
  <path:IMAGE.md>
  - image name: **FE**  build id 0xE4B711FE  role TAIL  state: withdrawn (coverage plan)  pdi md5 n/a
- image:0xE4B711FF |  | current | Image 0xE4B711FF [0xE4B711FF]
  <path:IMAGE.md>
  - image name: **FF**  build id 0xE4B711FF  role TAIL  state: LANDED build-host-M 11:33Z - HELD (superseded by FS)  pdi md5 c65579a6
- image:0xE4B71200 |  | current | Image 0xE4B71200
  <path:IMAGE.md>
  - image name: **HQ_CL4**  build id 0xE4B71200  role HEAD  state: LANDED 2026-10-04 11:36 build-vm = early copy already on lab-host (md5 5a4f219d [M] both ends); HOLD -> PASS (QM-19 manual): 5 'new' classes
- image:0xE4B71200_HQ_CL4h |  | current | Image 0xE4B71200_HQ_CL4h
  <path:IMAGE.md>
  - image name: **HQ_CL4h**  build id 0xE4B71200  role HEAD  state: LANDED 2026-10-04 12:38 build-vm = early copy 3d54898e [M]; SPARE (HQ_CL4 5a4f219d is the CL4 HEAD of record); same CL4 checker-artifac
- image:0xE4B71200_HQ_CL4s2 |  | current | Image 0xE4B71200_HQ_CL4s2
  <path:IMAGE.md>
  - image name: **HQ_CL4s2**  build id 0xE4B71200  role HEAD  state: LANDED 2026-10-05 18:09 EDT build-vm (wall 52,902 s, peak 110.8 GB vs reserved 97); PASS; RELAYED 18:12 FOUR_CARD/CL4 as HEAD SPARE + 
- image:0xE4B71201 |  | current | Image 0xE4B71201
  <path:IMAGE.md>
  - image name: **TQ_CL4**  build id 0xE4B71201  role TAIL  state: LANDED 2026-10-04 06:36; HOLD -> PASS (HQ tq_cl4_adj): 5 'new' classes = checker artifacts -> RELAYED lab-host pdi_inbox/FOUR_CARD/CL4 (
- image:0xE4B71201_TQ_CL4s2 |  | current | Image 0xE4B71201_TQ_CL4s2
  <path:IMAGE.md>
  - image name: **TQ_CL4s2**  build id 0xE4B71201  role TAIL  state: LANDED 2026-10-05 ~16:40 EDT build-vm; PASS (QM-20) - better-margin CL4 TAIL; RELAYED create-only lab-host FOUR_CARD/CL4 (md5 verified 
- image:0xE4B71202 |  | current | Image 0xE4B71202
  <path:IMAGE.md>
  - image name: **HQ_MP2h**  build id 0xE4B71202  role HEAD  state: LANDED 2026-10-04 13:12 build-host-A (PDI written 13:12, official 13:19); FAIL timing - not relayed  pdi md5 d44d4a3e
- image:0xE4B71203 |  | current | Image 0xE4B71203
  <path:IMAGE.md>
  - image name: **HX3TF**  build id 0xE4B71203  role HEAD  state: LANDED 2026-10-05 03:17 EDT build-vm (mesh EXIT rc=0 wall 54,461 s, measured peak 91.4 GB); TIMING PASS; HOLD - not relayed  pdi md5 ce8c0
- image:0xE4B71204 |  | current | Image 0xE4B71204
  <path:IMAGE.md>
  - image name: **HX3T0F**  build id 0xE4B71204  role HEAD  state: LANDED 2026-10-05 03:07 EDT build-vm (mesh EXIT rc=0 wall 65,603 s, measured peak 94.6 GB); TIMING FAIL; HOLD - not relayed  pdi md5 3d5
- image:0xE4B71205 |  | current | Image 0xE4B71205
  <path:IMAGE.md>
  - image name: **HX3T0FP**  build id 0xE4B71205  role HEAD  state: TERMINATED 2026-10-04 18:45 EDT (mesh EXIT rc=-15 SIGTERM after 6h42, measured peak 64.1 GB; pkg synth had passed); no PDI. Not stoppe
- image:0xE4B71206 |  | current | Image 0xE4B71206
  <path:IMAGE.md>
  - image name: **SG_HEAD**  build id 0xE4B71206  role HEAD  state: ROUTED FAIL [M] post_route_phys_opt WNS -0.515 / TNS -30,908 / WHS 0.000 (runme.log 2026-10-05 21:58 EDT); usr_2 KVI_FAN cone (QM-22 c
- image:0xE4B71206_SG_HEAD_v2 |  | current | Image 0xE4B71206_SG_HEAD_v2
  <path:IMAGE.md>
  - image name: **SG_HEAD_v2**  build id 0xE4B71206  role HEAD  state: STOPPED 2026-10-05 22:47 EDT (SIGTERM, not QM; never left the STEP4 heavy-lock wait); not requeued per HQ (xdc-only cannot close)  
- image:0xE4B71206_SG_HEAD_v3 |  | current | Image 0xE4B71206_SG_HEAD_v3
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3**  build id 0xE4B71206  role HEAD  state: BUILDING 2026-10-06 00:47 EDT mesh build-vm-1922644 peak 72 (QM-22, HQ GO)  pdi md5 n/a
- image:0xE4B71206_SG_HEAD_v3acr |  | current | Image 0xE4B71206_SG_HEAD_v3acr
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3acr**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-06 17:48 EDT (wall 61,263 s, peak 81.1 GB); FAIL hold WHS -0.016 (220 EP usr_1)  pdi md5 5af8733ed87b0d0b67421cdb
- image:0xE4B71206_SG_HEAD_v3ae |  | current | Image 0xE4B71206_SG_HEAD_v3ae
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3ae**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-06 11:05 EDT (wall 24,756 s, peak 74.7 GB); FAIL hold: WHS -0.016 (26 EP usr_1); usr_2 setup closed +0.075  pdi md
- image:0xE4B71206_SG_HEAD_v3dch |  | current | Image 0xE4B71206_SG_HEAD_v3dch
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3dch**  build id 0xE4B71206  role HEAD  state: RELAYED 2026-10-07 05:22 EDT create-only to SHAPE_GENERIC/PAIR2 as SPARE HEAD (NOTE_QM26_SG_HEAD_v3dch_b23e1330_SPARE.md: prefer
- image:0xE4B71206_SG_HEAD_v3dchx |  | current | Image 0xE4B71206_SG_HEAD_v3dchx
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3dchx**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-07 23:06 EDT (mesh build-vm-3045081 rc=0 wall 105,378 s peak 81.3 GB); FAIL timing (usr_1 -0.111, 10,407 EP) - no
- image:0xE4B71206_SG_HEAD_v3hm |  | current | Image 0xE4B71206_SG_HEAD_v3hm
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hm**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-06 17:57 EDT (wall 24,378 s, peak 70.9 GB); PASS; RELAYED 18:07 EDT as PAIR2 HEAD with SG_TAIL e403c6d4 to SHAPE_G
- image:0xE4B71206_SG_HEAD_v3hm30 |  | current | Image 0xE4B71206_SG_HEAD_v3hm30
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hm30**  build id 0xE4B71206  role HEAD  state: RELAYED 2026-10-07 18:59 EDT create-only to SHAPE_GENERIC/PAIR2 as SPARE, md5sum -c OK on lab-host (delayed from 13:45: lab-host down f
- image:0xE4B71206_SG_HEAD_v3hmd |  | current | Image 0xE4B71206_SG_HEAD_v3hmd
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hmd**  build id 0xE4B71206  role HEAD  state: RELAYED 2026-10-07 11:23 EDT create-only to SHAPE_GENERIC/PAIR2 as SPARE HEAD (NOTE_QM26_SG_HEAD_v3hmd_2f9d6d1f_SPARE.md: cdb3f6
- image:0xE4B71206_SG_HEAD_v3hmx |  | current | Image 0xE4B71206_SG_HEAD_v3hmx
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hmx**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-07 01:18 EDT (wall 25,816 s, peak 72.8 GB); PASS; RELAYED 01:29 EDT create-only to SHAPE_GENERIC/PAIR2 as alterna
- image:0xE4B71206_SG_HEAD_v3hmx30 |  | current | Image 0xE4B71206_SG_HEAD_v3hmx30
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hmx30**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-07 21:30 EDT (mesh build-host-A-2826540 rc=0 wall 38,526 s peak 63.2 GB); FAIL timing (usr_1 -0.055, 3,472 EP) - not 
- image:0xE4B71206_SG_HEAD_v3hmxd |  | current | Image 0xE4B71206_SG_HEAD_v3hmxd
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hmxd**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-07 12:25 EDT (mesh build-host-M-3709580 rc=0 wall 30,715 s peak 80.8 GB); PASS SPARE (0 new classes vs cdb3f63d); REL
- image:0xE4B71206_SG_HEAD_v3hmxvm |  | current | Image 0xE4B71206_SG_HEAD_v3hmxvm
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hmxvm**  build id 0xE4B71206  role HEAD  state: WITHDRAWN 2026-10-08 07:16:58 before admission (design lead: v3hmx wedges in S_KVL on silicon, F1040; norin is the replacement
- image:0xE4B71206_SG_HEAD_v3mg |  | current | Image 0xE4B71206_SG_HEAD_v3mg
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3mg**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-06 19:27 EDT (wall 31,914 s, peak 80.4 GB); FAIL setup -0.261 / hold -0.016  pdi md5 aef676f985f54540c10d95dcad6b8
- image:0xE4B71206_SG_HEAD_v3px |  | current | Image 0xE4B71206_SG_HEAD_v3px
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3px**  build id 0xE4B71206  role HEAD  state: LANDED 2026-10-06 21:56 EDT (wall 55,121 s, peak 80.3 GB); FAIL hold WHS -0.016 (111 EP usr_1, DSP-cascade signature)  pdi md5 19
- image:0xE4B71207 |  | current | Image 0xE4B71207
  <path:IMAGE.md>
  - image name: **SG_TAIL**  build id 0xE4B71207  role TAIL  state: RELAYED 2026-10-06 18:07 EDT as PAIR2 TAIL with SG_HEAD v3hm d8824e6b to SHAPE_GENERIC/PAIR2 (md5 OK on lab-host)  pdi md5 e403c6d439cd5c2
- image:0xE4B71207_SG_TAIL_hm |  | current | Image 0xE4B71207_SG_TAIL_hm
  <path:IMAGE.md>
  - image name: **SG_TAIL_hm**  build id 0xE4B71207  role TAIL  state: RELAYED 2026-10-07 11:22 EDT create-only to SHAPE_GENERIC/PAIR2 as SPARE TAIL (NOTE_QM26_SG_TAIL_hm_1973b51b_SPARE.md: e403c6d4 pre
- image:0xE4B71207_SG_TAIL_hmx |  | current | Image 0xE4B71207_SG_TAIL_hmx
  <path:IMAGE.md>
  - image name: **SG_TAIL_hmx**  build id 0xE4B71207  role TAIL  state: RELAYED 2026-10-07 18:59 EDT create-only to SHAPE_GENERIC/PAIR2 as recommended TAIL (NOTE_QM27_SG_TAIL_hmx_978ec0f9.md), md5sum -c
- image:0xE4B71207_SG_TAIL_px |  | current | Image 0xE4B71207_SG_TAIL_px
  <path:IMAGE.md>
  - image name: **SG_TAIL_px**  build id 0xE4B71207  role TAIL  state: LANDED 2026-10-07 11:35 EDT (wall 50,824 s, peak 62.3 GB); PASS, better margin than e403c6d4 -> RELAYED 11:36 EDT create-only to SH
- image:0xE4B71208 |  | current | Image 0xE4B71208
  <path:IMAGE.md>
  - image name: **SG_CL4H**  build id 0xE4B71208  role HEAD  state: LANDED 2026-10-05 19:20 EDT (verdicts 20:50); PASS at zero margin post_route_phys_opt WNS 0.000 / WHS 0.000, 0 failing EP (route_desig
- image:0xE4B71208_SG_CL4H_v2 |  | current | Image 0xE4B71208_SG_CL4H_v2
  <path:IMAGE.md>
  - image name: **SG_CL4H_v2**  build id 0xE4B71208  role HEAD  state: WITHDRAWN 2026-10-05 19:15 EDT before submit (HQ: xdc-only cannot close; awaiting SG_CL4H v4 = v5 RTL)  pdi md5 n/a
- image:0xE4B71208_SG_CL4H_v4 |  | current | Image 0xE4B71208_SG_CL4H_v4
  <path:IMAGE.md>
  - image name: **SG_CL4H_v4**  build id 0xE4B71208  role HEAD  state: LANDED PDI 2026-10-06 17:25 EDT (verdicts 18:11; async EVAC census still running); FAIL setup WNS -0.231 (27,173 EP usr_1, reset-fa
- image:0xE4B71208_SG_CL4H_v5 |  | current | Image 0xE4B71208_SG_CL4H_v5
  <path:IMAGE.md>
  - image name: **SG_CL4H_v5**  build id 0xE4B71208  role HEAD  state: LANDED PDI 2026-10-07 15:40Z (mesh EXIT 17:50Z rc=0 wall 68,056 s peak 81.6 GB); PASS; relay to SHAPE_GENERIC/FOUR_CARD staged, box
- image:0xE4B71208_SG_CL4H_v5b |  | current | Image 0xE4B71208_SG_CL4H_v5b
  <path:IMAGE.md>
  - image name: **SG_CL4H_v5b**  build id 0xE4B71208  role HEAD  state: RELAYED 2026-10-07 18:58 EDT create-only to SHAPE_GENERIC/FOUR_CARD as SPARE by shape_generic relay_when_up (RELAY_OK, md5 OK) + Q
- image:0xE4B71208_SG_CL4H_v5c |  | current | Image 0xE4B71208_SG_CL4H_v5c
  <path:IMAGE.md>
  - image name: **SG_CL4H_v5c**  build id 0xE4B71208  role HEAD  state: KILLED rc=-15 SIGTERM 2026-10-07 06:50Z by unknown party after 2.2 h STEP4 heavy-lock wait behind v5b (shared slot_gCP1Q_dc_v4 loc
- image:0xE4B71208_SG_CL4H_v5c2 |  | current | Image 0xE4B71208_SG_CL4H_v5c2
  <path:IMAGE.md>
  - image name: **SG_CL4H_v5c2**  build id 0xE4B71208  role HEAD  state: FAILED 2026-10-08 00:07 EDT (mesh build-vm-892728 rc=1, wall 26,012 s, peak 85.7 GB): no PDI - place.pre.tcl DCGRP stanza Tcl erro
- image:0xE4B71208_SG_CL4H_v5c3 |  | current | Image 0xE4B71208_SG_CL4H_v5c3
  <path:IMAGE.md>
  - image name: **SG_CL4H_v5c3**  build id 0xE4B71208  role HEAD  state: PASS SPARE QM-34; RELAYED 2026-10-09 16:19 EDT create-only SHAPE_GENERIC/FOUR_CARD + NOTE_QM34 do-not-flash (md5 -c OK); v5 07849
- image:0xE4B71209 |  | current | Image 0xE4B71209
  <path:IMAGE.md>
  - image name: **SG_CL4T**  build id 0xE4B71209  role TAIL  state: LANDED 2026-10-06 07:28 EDT (mesh EXIT rc=0 wall 52,682 s peak 87.5 GB); PASS ~zero margin; RELAYED 2026-10-06 08:22 EDT with SG_CL4H 
- image:0xE4B71209_SG_CL4T_v2 |  | current | Image 0xE4B71209_SG_CL4T_v2
  <path:IMAGE.md>
  - image name: **SG_CL4T_v2**  build id 0xE4B71209  role TAIL  state: QUEUED 2026-10-05 18:45 EDT: submit_v3_when_free.sh (pid 1547889) submits peak 100 when HX2QRs2 or SG_CL4T v1 exits  pdi md5 n/a
- image:0xE4B71209_SG_CL4T_v3 |  | current | Image 0xE4B71209_SG_CL4T_v3
  <path:IMAGE.md>
  - image name: **SG_CL4T_v3**  build id 0xE4B71209  role TAIL  state: LANDED 2026-10-06 12:56 EDT (wall 44,252 s, peak 97.9 GB); PASS zero margin; RELAYED 12:58 EDT to SHAPE_GENERIC/FOUR_CARD (NOTE_QM2
- image:0xE4B7120A |  | current | Image 0xE4B7120A
  <path:IMAGE.md>
  - image name: **HBMS_A400**  build id 0xE4B7120A  role SINGLE  state: STAGED 2026-10-07 05:52 EDT, not submitted (as before: relaunch when a host frees)  pdi md5 n/a
- image:0xE4B7120B |  | current | Image 0xE4B7120B
  <path:IMAGE.md>
  - image name: **HBMS_C_MD4**  build id 0xE4B7120B  role SINGLE  state: LANDED VERDICT FAIL (WNS -5.432 WHS 0.000); not relayed  pdi md5 n/a
- image:0xE4B7120C |  | current | Image 0xE4B7120C
  <path:IMAGE.md>
  - image name: **HBMS_C_MD2**  build id 0xE4B7120C  role SINGLE  state: LANDED VERDICT FAIL (WNS -2.910 WHS 0.000); not relayed  pdi md5 n/a
- image:0xE4B7120D |  | current | Image 0xE4B7120D
  <path:IMAGE.md>
  - image name: **HBMS_C_MD4_333**  build id 0xE4B7120D  role SINGLE  state: LANDED VERDICT FAIL (WNS -2.685 WHS 0.000); not relayed  pdi md5 n/a
- image:0xE4B7120E |  | current | Image 0xE4B7120E
  <path:IMAGE.md>
  - image name: **PA_HEAD_P34**  build id 0xE4B7120E  role HEAD  state: PDI 2026-10-08 05:05 EDT + gates 05:53 (mesh job build-vm-499627 still open for async EVAC tooth) -> QM-28 PASS -> RELAYED 06:46 POR
- image:0xE4B7120E_PA_HEAD_P34hm |  | current | Image 0xE4B7120E_PA_HEAD_P34hm
  <path:IMAGE.md>
  - image name: **PA_HEAD_P34hm**  build id 0xE4B7120E  role HEAD  state: LANDED 2026-10-08 17:31 EDT (PDI 16:41); QM-29 HOLD: NEW EVAC-SLL10 FAIL (SLR1<->SLR0 SLL 90.64% > 90%) + G1/G5 census rows; not
- image:0xE4B7120F |  | current | Image 0xE4B7120F
  <path:IMAGE.md>
  - image name: **PA_TAIL_P34**  build id 0xE4B7120F  role TAIL  state: LANDED 2026-10-08 04:23 EDT (mesh build-vm-720777 rc=0 wall 49,719 s peak 96.7 GB) -> RELAYED 04:27 create-only PORT_AGNOSTIC + FOU
- image:0xE4B7120F_PA_TAIL_P34hm |  | current | Image 0xE4B7120F_PA_TAIL_P34hm
  <path:IMAGE.md>
  - image name: **PA_TAIL_P34hm**  build id 0xE4B7120F  role TAIL  state: LANDED 2026-10-08 01:45 EDT (mesh build-vm-3697529 rc=0 wall 77,194 s peak 96.9 GB) -> RELAYED 01:5x create-only PORT_AGNOSTIC + 
- image:0xE4B71210 |  | current | Image 0xE4B71210
  <path:IMAGE.md>
  - image name: **HBMS_L**  build id 0xE4B71210  role SINGLE  state: STAGED 2026-10-07 07:22 EDT, not submitted (no build host > 65% real free RAM: build-vm 54%, build-vm 48%, build-vm 44%, build-host-M 39%, asg
- image:0xE4B71210_HBMS_L_r4e |  | current | Image 0xE4B71210_HBMS_L_r4e
  <path:IMAGE.md>
  - image name: **HBMS_L_r4e**  build id 0xE4B71210  role HEAD  state: SUPERSEDED-TIMING 2026-10-07 14:13 EDT: whole-core OOC WNS -0.497 ns / 1,229 failing EP [M] (F0960; only hbml_central closed +0.049
- image:0xE4B71211 |  | current | Image 0xE4B71211
  <path:IMAGE.md>
  - image name: **HBMS_KV**  build id 0xE4B71211  role SINGLE  state: QUEUED 2026-10-07 14:03 EDT mesh waiter build-vm-686071-396192217 (hbms_build_L peak 48; build-vm real free 176/251 GB = 70% [M]; behi
- image:0xE4B71211_HBMS_L_r5h |  | current | Image 0xE4B71211_HBMS_L_r5h
  <path:IMAGE.md>
  - image name: **HBMS_L_r5h**  build id 0xE4B71211  role HEAD  state: SUPERSEDED 2026-10-07 15:27 EDT: waiter build-vm-686071 WITHDRAWN via withdraw_waiter.sh (never admitted, never built); replaced by 
- image:0xE4B71212 |  | current | Image 0xE4B71212
  <path:IMAGE.md>
  - image name: **D2P34_HEAD**  build id 0xE4B71212  role HEAD  state: LANDED 2026-10-08 08:31 EDT (PDI written; mesh build-host-A-256418 open for async EVAC tooth); QM-29 hand verdict PASS (NEW=5, 0 real) ->
- image:0xE4B71212_D2P34_HEADhm |  | current | Image 0xE4B71212_D2P34_HEADhm
  <path:IMAGE.md>
  - image name: **D2P34_HEADhm**  build id 0xE4B71212  role HEAD  state: LANDED PDI 2026-10-08 10:43 EDT (mesh build-vm-102346 still in post-PDI gates); QM-29 hand verdict PASS (NEW=5, 0 real) -> RELAYED
- image:0xE4B71212_D2P34_HEADhmx |  | current | Image 0xE4B71212_D2P34_HEADhmx
  <path:IMAGE.md>
  - image name: **D2P34_HEADhmx**  build id 0xE4B71212  role HEAD  state: LANDED 2026-10-09 03:22 EDT PDI (build-vm); QM-33 verdict FAIL: setup WNS -0.474, 60,823 failing EP [M]; NOT relayed  pdi md5 5d15
- image:0xE4B71212_D2P34_HEADmd |  | current | Image 0xE4B71212_D2P34_HEADmd
  <path:IMAGE.md>
  - image name: **D2P34_HEADmd**  build id 0xE4B71212  role HEAD  state: BLOCKED 2026-10-08 03:37: pipeline STEP0-pre disk guard needs 60 GB on /mnt/data2, has 36 GB -> rc=1 at start, nothing built; res
- image:0xE4B71213 |  | current | Image 0xE4B71213
  <path:IMAGE.md>
  - image name: **D2P34_TAIL**  build id 0xE4B71213  role TAIL  state: LANDED 2026-10-08 18:46 EDT (GV 19:00, mesh build-vm-1706638 rc=0 peak 66.1 GB); QM-29 hand verdict PASS (NEW=5, 0 real) -> RELAYED 
- image:0xE4B71213_D2P34_TAILhm |  | current | Image 0xE4B71213_D2P34_TAILhm
  <path:IMAGE.md>
  - image name: **D2P34_TAILhm**  build id 0xE4B71213  role TAIL  state: LANDED 2026-10-08 06:12 EDT (mesh build-vm-631265 rc=0 wall 61,410 s peak 66.3 GB); watcher HOLD NEW=5 -> QM-28 hand verdict PASS 
- image:0xE4B71214 |  | current | Image 0xE4B71214
  <path:IMAGE.md>
  - image name: **HBMS_MS**  build id 0xE4B71214  role SINGLE  state: STAGED 2026-10-07, not submitted (build-vm real free 57% < 65%): /shared/build/hbmsMS1214r5 (mk_stage_ms5.sh, launch.sh). Strict super
- image:0xE4B71214_HBMS_MS_r5h |  | current | Image 0xE4B71214_HBMS_MS_r5h
  <path:IMAGE.md>
  - image name: **HBMS_MS_r5h**  build id 0xE4B71214  role HEAD  state: BUILDING 2026-10-07 19:33 EDT: moved from build-vm (waiter build-vm-685478 withdrawn, never admitted) to build-host-M ~/hbms_build/hbmsMS12
- image:0xE4B71215 |  | current | Image 0xE4B71215
  <path:IMAGE.md>
  - image name: **HBMS_KVW_r6i**  build id 0xE4B71215  role SINGLE  state: LANDED VERDICT FAIL (WNS -3.122 WHS 0.000); not relayed  pdi md5 42b33529ee7785344a7f70c733d82a58
- image:0xE4B71216 |  | current | Image 0xE4B71216
  <path:IMAGE.md>
  - image name: **HBMS_SPEC_r7vl**  build id 0xE4B71216  role SINGLE  state: QUEUED 2026-10-07 21:28 EDT mesh waiter build-vm-1293615-422915189 (hbms_build_L, admitted against 65.2 GB learned peak; launc
- image:0xE4B71217 |  | current | Image 0xE4B71217
  <path:IMAGE.md>
  - image name: **HBMS_SPEC6_np16_300**  build id 0xE4B71217  role SINGLE  state: BUILDING 2026-10-08 03:37 EDT mesh id build-host-M-839410-444276076 (hbms6_build_np16, slot ~/hbms_build/hbmsSPEC6_1217np16);
- image:0xE4B71218 |  | current | Image 0xE4B71218
  <path:IMAGE.md>
  - image name: **HBMS_SPEC6_np16slr_300**  build id 0xE4B71218  role SINGLE  state: QUEUED 2026-10-08 03:37 EDT mesh waiter build-host-A-483103-445031523 (hbms6_build_np16, slot /shared/build/hbmsSPEC6_1218n
- image:0xE4B71219 |  | current | Image 0xE4B71219
  <path:IMAGE.md>
  - image name: **SG_HEAD_v3hmx_norin**  build id 0xE4B71219  role HEAD  state: LANDED 2026-10-09 (PDI on build-vm); QM-33 verdict FAIL: setup WNS -0.252, 21,587 failing EP [M]; NOT relayed  pdi md5 3fe1
- image:0xE4B7121A |  | current | Image 0xE4B7121A
  <path:IMAGE.md>
  - image name: **HBMS_SPEC7_np16_300**  build id 0xE4B7121A  role SINGLE  state: QUEUED 2026-10-08 05:31 EDT mesh waiter build-host-M-973262-451892123 (hbms7_build_np16 peak 50, slot ~/hbms_build/hbmsSPEC7_
- image:0xE4B7121B |  | current | Image 0xE4B7121B
  <path:IMAGE.md>
  - image name: **3S_A_ATTN**  build id 0xE4B7121B  role HEAD  state: QUEUED 2026-10-08 13:24 EDT mesh build-vm-1992499-465899208 peak 80 (after hard re-gate GO: +8.9..+10.8% raw [D], regate_final.out)  
- image:0xE4B7121B_3S_A_ATTN_s3LA |  | current | Image 0xE4B7121B_3S_A_ATTN_s3LA
  <path:IMAGE.md>
  - image name: **3S_A_ATTN_s3LA**  build id 0xE4B7121B  role HEAD  state: LANDED 2026-10-09 09:29 EDT (PIPELINE COMPLETE 09:42); QM-33 hand PASS (0 real NEW vs 974b37a5) -> RELAYED 09:43 THREE_CARD/V3,
- image:0xE4B7121B_3S_A_s3LAm3md |  | current | Image 0xE4B7121B_3S_A_s3LAm3md
  <path:IMAGE.md>
  - image name: **3S_A_s3LAm3md**  build id 0xE4B7121B  role HEAD  state: PASS SPARE QM-36; RELAYED 2026-10-10 13:50 EDT create-only THREE_CARD/V3 + NOTE do-not-flash (md5 -c OK); equal margin to f33ea5
- image:0xE4B7121B_3S_A_s3LAm4 |  | current | Image 0xE4B7121B_3S_A_s3LAm4
  <path:IMAGE.md>
  - image name: **3S_A_s3LAm4**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: PASS at zero margin QM-38 (usr_1 0.000, 0 new classes); no upgrade over f33ea554; NOT relayed  pdi md5 e97b201feca
- image:0xE4B7121B_3S_A_s3LAm5md |  | current | Image 0xE4B7121B_3S_A_s3LAm5md
  <path:IMAGE.md>
  - image name: **3S_A_s3LAm5md**  build id 0xE4B7121B  role HEAD  state: WITHDRAWN QM-36 14:25 EDT (pre-admission waiter, withdraw_waiter.sh GONE): HQ moved build-host-M to s3LAu3 (directive seeds all land 
- image:0xE4B7121B_3S_A_s3LAmd |  | current | Image 0xE4B7121B_3S_A_s3LAmd
  <path:IMAGE.md>
  - image name: **3S_A_s3LAmd**  build id 0xE4B7121B  role HEAD  state: FAIL QM-34: WNS -0.001, 187 failing EP [M]; PDI d2ee404c not relayed  pdi md5 d2ee404c
- image:0xE4B7121B_3S_A_s3LAr1c |  | current | Image 0xE4B7121B_3S_A_s3LAr1c
  <path:IMAGE.md>
  - image name: **3S_A_s3LAr1c**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: SPARE-1 card A (HQ): R1-retime RTL, true usr_1 +0.201 [D]; RELAYED 2026-10-11 21:16 EDT THREE_CARD/V3 (pdi+MD5+GV
- image:0xE4B7121B_3S_A_s3LAu3 |  | current | Image 0xE4B7121B_3S_A_s3LAu3
  <path:IMAGE.md>
  - image name: **3S_A_s3LAu3**  build id 0xE4B7121B  role HEAD  state: STOPPED by QM-36 17:22 EDT (own job; HQ order): idle 35+ min in STEP4 heavy_impl.lock wait behind s3LAm4 (BR085 shared-lock bug); 
- image:0xE4B7121B_3S_A_s3LAu3b |  | current | Image 0xE4B7121B_3S_A_s3LAu3b
  <path:IMAGE.md>
  - image name: **3S_A_s3LAu3b**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: SPARE card-A image (HQ); primary = s3LAu4md 9766f14d  pdi md5 0421a8ff9c17b080a2cd0949396c52ef
- image:0xE4B7121B_3S_A_s3LAu4md |  | current | Image 0xE4B7121B_3S_A_s3LAu4md
  <path:IMAGE.md>
  - image name: **3S_A_s3LAu4md**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: PRIMARY card-A image (HQ); s3LAu3b 0421a8ff = SPARE; f33ea554 superseded for next flash opportunity; ADDENDUM_QM
- image:0xE4B7121B_3S_A_s3LAu6 |  | current | Image 0xE4B7121B_3S_A_s3LAu6
  <path:IMAGE.md>
  - image name: **3S_A_s3LAu6**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: PASS but no upgrade QM-38: true usr_1 +0.150 [D] < primary s3LAu4md +0.250; 0 new classes vs f33ea554; NOT relayed
- image:0xE4B7121B_s3LAm2 |  | current | Image 0xE4B7121B_s3LAm2
  <path:IMAGE.md>
  - image name: **s3LAm2**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: FAIL QM-37: post-route WNS -0.276 (usr_1) / 38815 failing / WHS 0.000; PDI 10a59b1e not relayed (F1204)  pdi md5 10a59b
- image:0xE4B7121B_s3LAr1c |  | current | Image 0xE4B7121B_s3LAr1c
  <path:IMAGE.md>
  - image name: **s3LAr1c**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: BUILDING (card_a_pace 2026-10-11): mesh build-vm-3781504 ADMITTED before the HQ hold arrived, Vivado running [M 04:3x];
- image:0xE4B7121B_s3LAr1u |  | current | Image 0xE4B7121B_s3LAr1u
  <path:IMAGE.md>
  - image name: **s3LAr1u**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: STAGED-HELD (card_a_pace 2026-10-11): pre-admission waiter build-vm-3187420 WITHDRAWN 04:06 (withdraw_waiter GONE) per 
- image:0xE4B7121B_s3LAu5 |  | current | Image 0xE4B7121B_s3LAu5
  <path:IMAGE.md>
  - image name: **s3LAu5**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state:  HOLD 0.020) + impl-only UU usr_0 0.300 / usr_1 0.200 / usr_2 0.080 (xdc 0f998512; run fc9acc7c; per-slot lock slot_s3L
- image:0xE4B7121B_s3LAu6 |  | current | Image 0xE4B7121B_s3LAu6
  <path:IMAGE.md>
  - image name: **s3LAu6**  build id 0xE4B7121B  role HEAD (3STAGE card A)  state: BUILDING QM-37: mesh build-vm-3553303 class pdi_build_head_s3LAu6 peak 86 ADMITTED 23:13 EDT; lint v4+v5 PASS. Supersede 
- image:0xE4B7121C |  | current | Image 0xE4B7121C
  <path:IMAGE.md>
  - image name: **3S_B_MLP**  build id 0xE4B7121C  role MID  state: QUEUED 2026-10-08 13:49 EDT SLOT s3LBmd2 build-host-M-1435090-481756087 peak 80 (waits on stale s3LBmd reservation ~77 GB, QM to release)  
- image:0xE4B7121C_3S_B_MID_s3LBmd |  | current | Image 0xE4B7121C_3S_B_MID_s3LBmd
  <path:IMAGE.md>
  - image name: **3S_B_MID_s3LBmd**  build id 0xE4B7121C  role HEAD  state: BUILDING: ADMITTED 2026-10-08 12:01:46 EDT build-host-M-1325207 (moved from build-vm waiter 2014104, withdrawn 12:02:48); STEP1 pack
- image:0xE4B7121C_3S_B_MID_s3LBmd2 |  | current | Image 0xE4B7121C_3S_B_MID_s3LBmd2
  <path:IMAGE.md>
  - image name: **3S_B_MID_s3LBmd2**  build id 0xE4B7121C  role MID  state: LANDED 2026-10-09 (mesh EXIT 01:46:48 rc=0); QM-33 hand PASS (NEW 8 vs 974b37a5: 7 artifacts/expected, EVAC-SLL10 91.97% soft)
- image:0xE4B7121C_3S_B_dc15487f |  | current | Image 0xE4B7121C_3S_B_dc15487f
  <path:IMAGE.md>
  - image name: **3S_B_dc15487f**  build id 0xE4B7121C  role MID (3STAGE card B)  state: SPARE card-B image (was primary; on card 61); superseded by s3LBu1 dd7ea3d9  pdi md5 dc15487f
- image:0xE4B7121C_3S_B_s3LBu1 |  | current | Image 0xE4B7121C_3S_B_s3LBu1
  <path:IMAGE.md>
  - image name: **3S_B_s3LBu1**  build id 0xE4B7121C  role MID (3STAGE card B)  state: PRIMARY card-B image (HQ); dc15487f = SPARE; ADDENDUM_QM38_CARD_B_PRIMARY.md relayed THREE_CARD/V3 (md5 -c OK); box
- image:0xE4B7121C_s3LBu1 |  | current | Image 0xE4B7121C_s3LBu1
  <path:IMAGE.md>
  - image name: **s3LBu1**  build id 0xE4B7121C  role MID (3STAGE card B)  state: BUILDING QM-37: mesh build-vm-3732493 class pdi_build_head_s3LBu1 peak 86 ADMITTED 02:59 EDT; lint v4+v5+MODREF PASS. Sup
- image:0xE4B7121D |  | current | Image 0xE4B7121D
  <path:IMAGE.md>
  - image name: **3S_C_LMTREE**  build id 0xE4B7121D  role TAIL  state: QUEUED 2026-10-08 13:38 EDT mesh build-vm-2038401-466712470 peak 72 (after hard re-gate GO; G0 strap-0 + RED2 xsim still running, RT
- image:0xE4B7121D_3S_C_LMTREE_s3LCr |  | current | Image 0xE4B7121D_3S_C_LMTREE_s3LCr
  <path:IMAGE.md>
  - image name: **3S_C_LMTREE_s3LCr**  build id 0xE4B7121D  role TAIL  state: LANDED 2026-10-09 07:31 EDT; QM-33 hand PASS -> RELAYED 07:42 THREE_CARD/V3, md5 OK on lab-host, NOTE: C -> 62, do not flash unt
- image:0xE4B7121D_3S_C_LMTREEcg |  | current | Image 0xE4B7121D_3S_C_LMTREEcg
  <path:IMAGE.md>
  - image name: **3S_C_LMTREEcg**  build id 0xE4B7121D  role TAIL  state: QUEUED 2026-10-08 22:05 EDT build-vm-2354309 (waits norin exit)  pdi md5 n/a
- image:0xE4B7121D_3S_C_s3LCcg2 |  | current | Image 0xE4B7121D_3S_C_s3LCcg2
  <path:IMAGE.md>
  - image name: **3S_C_s3LCcg2**  build id 0xE4B7121D  role TAIL  state: PASS SPARE QM-35; RELAYED 2026-10-09 22:30 EDT create-only THREE_CARD/V3 + NOTE do-not-flash (md5 -c OK)  pdi md5 ca4a2a5c
- image:0xE4B7121E |  | current | Image 0xE4B7121E
  <path:IMAGE.md>
  - image name: **3S_C_LMTREE80**  build id 0xE4B7121E  role TAIL  state: RESERVED 2026-10-08 (build only after the 3-card chain re-gate)  pdi md5 n/a
- image:0xE4B7121E_3S_C_LMTREE100 |  | current | Image 0xE4B7121E_3S_C_LMTREE100
  <path:IMAGE.md>
  - image name: **3S_C_LMTREE100**  build id 0xE4B7121E  role TAIL  state: RE-QUEUED 2026-10-08 12:49 EDT mesh build-vm-2162566-478147855 peak 72; old waiter build-vm-2122485 withdrawn (withdraw_waiter.sh G
- image:0xE4B7121E_s3LCh |  | current | Image 0xE4B7121E_s3LCh
  <path:IMAGE.md>
  - image name: **s3LCh**  build id 0xE4B7121E  role TAIL (3STAGE card C @100)  state: FAIL QM-37: post-route WNS -0.134 / 3559 failing / WHS 0.000; PDI 25cfd396 not relayed (F1203)  pdi md5 25cfd39613c
- image:0xE4B7121E_s3LCh3 |  | current | Image 0xE4B7121E_s3LCh3
  <path:IMAGE.md>
  - image name: **s3LCh3**  build id 0xE4B7121E  role TAIL (3STAGE card C @100)  state: FAIL QM-37: post-route WNS -0.193 / 7155 failing / WHS 0.000; PDI 808bf686 not relayed (F1202)  pdi md5 808bf686fa
- image:0xE4B7121F |  | current | Image 0xE4B7121F
  <path:IMAGE.md>
  - image name: **3S_C_LMTREE66**  build id 0xE4B7121F  role TAIL  state: BUILDING 2026-10-08 22:03 EDT build-vm-1664804  pdi md5 n/a
- image:0xE4B7121F_3S_C_s3LC66b |  | current | Image 0xE4B7121F_3S_C_s3LC66b
  <path:IMAGE.md>
  - image name: **3S_C_s3LC66b**  build id 0xE4B7121F  role TAIL  state: PASS SPARE QM-34; RELAYED 2026-10-09 18:46 EDT create-only THREE_CARD/V3 + NOTE do-not-flash (md5 -c OK)  pdi md5 720937ec
- image:0xE4B71230 |  | current | Image 0xE4B71230
  <path:IMAGE.md>
  - image name: **HBMS_SLR_base_300**  build id 0xE4B71230  role HBMS  state: BUILDING 2026-10-08 13:09 EDT mesh build-host-M-1384845 peak 57  pdi md5 n/a
- image:0xE4B71231 |  | current | Image 0xE4B71231
  <path:IMAGE.md>
  - image name: **HBMS_SLR_s21_300**  build id 0xE4B71231  role HBMS  state: QUEUED 2026-10-08 13:09 EDT mesh build-host-M-1391365 peak 57  pdi md5 n/a
- image:0xE4B71231_HBMS_SLR1231s21v3_300 |  | current | Image 0xE4B71231_HBMS_SLR1231s21v3_300
  <path:IMAGE.md>
  - image name: **HBMS_SLR1231s21v3_300**  build id 0xE4B71231  role HBMS  state: EXIT 2026-10-09 07:12 EDT rc=5; QM-33 FAIL: WNS -0.576, 194,837 failing EP [M]; not relayed  pdi md5 n/a
- image:0xE4B71232 |  | current | Image 0xE4B71232
  <path:IMAGE.md>
  - image name: **HBMS_SLR_s21_250**  build id 0xE4B71232  role HBMS  state: QUEUED 2026-10-08 build-vm-2178278 peak 57  pdi md5 n/a
- image:0xE4B71232_HBMS_SLR1232s21v2_250 |  | current | Image 0xE4B71232_HBMS_SLR1232s21v2_250
  <path:IMAGE.md>
  - image name: **HBMS_SLR1232s21v2_250**  build id 0xE4B71232  role HBMS  state: EXIT 2026-10-09 12:13 EDT rc=0; QM-34 FAIL: WNS -0.058, 2,566 failing EP [M]; PDI cf22b3e5 not relayed  pdi md5 cf22b3e5
- image:0xE4B71233 |  | current | Image 0xE4B71233
  <path:IMAGE.md>
  - image name: **HBMS_SLR_np32s6_200**  build id 0xE4B71233  role HBMS  state: BUILDING impl_1 (phys_opt) 2026-10-09 11:57 EDT mesh build-vm-2764641 (resume, synth_1 kept)  pdi md5 n/a
- image:0xE4B71233_HBMS_SLR1233np32s6_200 |  | current | Image 0xE4B71233_HBMS_SLR1233np32s6_200
  <path:IMAGE.md>
  - image name: **HBMS_SLR1233np32s6_200**  build id 0xE4B71233  role HBMS  state: PASS QM-34; RELAYED 2026-10-09 14:52 EDT create-only pdi_inbox/HBMS/SLR1233np32s6_200 + NOTE do-not-flash (md5 -c OK on
- image:0xE4B71233_HBMS_SLR1233np32s6e_200 |  | current | Image 0xE4B71233_HBMS_SLR1233np32s6e_200
  <path:IMAGE.md>
  - image name: **HBMS_SLR1233np32s6e_200**  build id 0xE4B71233  role HBMS  state: PASS SPARE QM-36; RELAYED 2026-10-10 14:47 EDT create-only pdi_inbox/HBMS/SLR1233np32s6e_200 + NOTE do-not-flash (md5 
- image:0xE4B71233_HBMS_SLR_np32s6e_200 |  | current | Image 0xE4B71233_HBMS_SLR_np32s6e_200
  <path:IMAGE.md>
  - image name: **HBMS_SLR_np32s6e_200**  build id 0xE4B71233  role HBMS  state: BUILDING mesh build-vm-2612484 hbms7_build_np32 peak 72 slot /shared/build/hbmsSLR_1233np32s6e_200 (QM-35 2026-10-10 06:02
- image:0xE4B71234 |  | current | Image 0xE4B71234
  <path:IMAGE.md>
  - image name: **HBMS_SLR_np32s6_250**  build id 0xE4B71234  role HBMS  state: PASS QM-35; RELAYED 2026-10-10 02:59 EDT create-only pdi_inbox/HBMS/SLR1234np32s6_250 + NOTE do-not-flash (md5 -c OK)  pdi
- image:0xE4B71235 |  | current | Image 0xE4B71235
  <path:IMAGE.md>
  - image name: **HBMS_SLR_np32s6_225**  build id 0xE4B71235  role HBMS  state: PASS QM-35; RELAYED 2026-10-10 06:06 EDT create-only pdi_inbox/HBMS/SLR1235np32s6_225 + NOTE do-not-flash (md5 -c OK)  pdi
- image:0xE4B71235_HBMS_SLR1235np32s6u_225 |  | current | Image 0xE4B71235_HBMS_SLR1235np32s6u_225
  <path:IMAGE.md>
  - image name: **HBMS_SLR1235np32s6u_225**  build id 0xE4B71235  role HBMS  state: PREFERRED NP32 @225 image (HQ); bb7097af = SPARE; ADDENDUM_QM36_PREFERRED_0x1235.md relayed to both HBMS folders 19:4x
- image:0xE4B71236 |  | current | Image 0xE4B71236
  <path:IMAGE.md>
  - image name: **HBMS_SWP1236A_225**  build id 0xE4B71236  role HBMS  state: PASS QM-35 (zero margin); RELAYED 2026-10-10 03:31 EDT create-only pdi_inbox/HBMS/SWP1236A_225 + NOTE do-not-flash (md5 -c O
- image:0xE4B7123B |  | current | Image 0xE4B7123B
  <path:IMAGE.md>
  - image name: **KVIFIX_A_kr1**  build id 0xE4B7123B  role Track1 record image (3-card drafter)  state: QM FAIL 2026-10-16 (margin)  pdi md5 b29db8b4
- image:0xE4B7123B_KVIFIX_A_kf4_laptop |  | current | Image 0xE4B7123B_KVIFIX_A_kf4_laptop
  <path:IMAGE.md>
  - image name: **KVIFIX_A_kf4_laptop**  build id 0xE4B7123B  role Track1 3-card head (laptop VM seed)  state: QM FAIL 2026-10-20 (timing; harvested from laptop)  pdi md5 b56ba1d0
- image:0xE4B7123B_KVIFIX_A_kr1b |  | current | Image 0xE4B7123B_KVIFIX_A_kr1b
  <path:IMAGE.md>
  - image name: **KVIFIX_A_kr1b**  build id 0xE4B7123B  role Track1 3-card head (insurance HOLD 0.035)  state: QM FAIL 2026-10-16 (margin)  pdi md5 cf038bb3
- image:0xE4B71240 |  | current | Image 0xE4B71240
  <path:IMAGE.md>
  - image name: **HBMS_SWP_A**  build id 0xE4B71240  role HBMS  state: LANDED PASS 2026-10-08 22:10 EDT; relayed pdi_inbox/HBMS/SWP1240A_200 (do not flash until go)  pdi md5 3ef124a4f34b3a7c46aff601c87c
- image:0xE4B71240_HBMS_SWP1240A_200 |  | current | Image 0xE4B71240_HBMS_SWP1240A_200
  <path:IMAGE.md>
  - image name: **HBMS_SWP1240A_200**  build id 0xE4B71240  role HBMS  state: SPARE (HQ 2026-10-11 02:5x: preferred [redacted-addr] = b97b06fd)  pdi md5 3ef124a4f34b3a7c46aff601c87c69b9
- image:0xE4B71240_HBMS_SWP1240As2_200 |  | current | Image 0xE4B71240_HBMS_SWP1240As2_200
  <path:IMAGE.md>
  - image name: **HBMS_SWP1240As2_200**  build id 0xE4B71240  role HBMS  state: PASS SPARE QM-35; RELAYED 2026-10-10 06:01 EDT create-only pdi_inbox/HBMS/SWP1240As2_200 + NOTE do-not-flash (md5 -c OK)  
- image:0xE4B71240_HBMS_SWP1240Au_200 |  | current | Image 0xE4B71240_HBMS_SWP1240Au_200
  <path:IMAGE.md>
  - image name: **HBMS_SWP1240Au_200**  build id 0xE4B71240  role HBMS  state: PREFERRED NP16 @200 image (HQ); 3ef124a4 = SPARE; ADDENDUM_QM37_PREFERRED_0x1240.md relayed to SWP1240A_200 + SWP1240Au_200
- image:0xE4B71241 |  | current | Image 0xE4B71241
  <path:IMAGE.md>
  - image name: **HBMS_SWP_B**  build id 0xE4B71241  role HBMS  state: QUEUED 2026-10-08 13:17 EDT mesh (hbms_sweep/launch_*.sh)  pdi md5 n/a
- image:0xE4B71241_HBMS_SWP1241Bs2_250 |  | current | Image 0xE4B71241_HBMS_SWP1241Bs2_250
  <path:IMAGE.md>
  - image name: **HBMS_SWP1241Bs2_250**  build id 0xE4B71241  role HBMS  state: PASS SPARE QM-35; RELAYED 2026-10-10 00:04 EDT create-only pdi_inbox/HBMS/SWP1241Bs2_250 + NOTE do-not-flash (md5 -c OK)  
- image:0xE4B71241_HBMS_SWP1241Bv2_250 |  | current | Image 0xE4B71241_HBMS_SWP1241Bv2_250
  <path:IMAGE.md>
  - image name: **HBMS_SWP1241Bv2_250**  build id 0xE4B71241  role HBMS  state: PASS QM-34; RELAYED 2026-10-09 16:23 EDT create-only pdi_inbox/HBMS/SWP1241Bv2_250 + NOTE do-not-flash (md5 -c OK)  pdi md
- image:0xE4B71242 |  | current | Image 0xE4B71242
  <path:IMAGE.md>
  - image name: **HBMS_SWP_C**  build id 0xE4B71242  role HBMS  state: QUEUED 2026-10-08 13:17 EDT mesh (hbms_sweep/launch_*.sh)  pdi md5 n/a
- image:0xE4B71242_HBMS_SWP1242Cm_200 |  | current | Image 0xE4B71242_HBMS_SWP1242Cm_200
  <path:IMAGE.md>
  - image name: **HBMS_SWP1242Cm_200**  build id 0xE4B71242  role HBMS  state: LANDED 2026-10-09 04:49 EDT (mesh rc=0); QM-33 PASS (adjudicate_hbms_r2) -> RELAYED 04:53 pdi_inbox/HBMS/SWP1242Cm_200, md5
- image:0xE4B71243 |  | current | Image 0xE4B71243
  <path:IMAGE.md>
  - image name: **HBMS_SWP_D**  build id 0xE4B71243  role HBMS  state: QUEUED 2026-10-08 13:17 EDT mesh (hbms_sweep/launch_*.sh)  pdi md5 n/a
- image:0xE4B71250 |  | current | Image 0xE4B71250
  <path:IMAGE.md>
  - image name: **R8C4_0x1250_b1**  build id 0xE4B71250  role RING8 C4 one-image (lab-host build)  state: PASS QM-38 (true hbms_clk +0.092 [D]; 0 new FAIL classes vs b334c397; Tier-E 18/0/1, kit-v2 rows 7/7
- image:0xE4B71251 |  | current | Image 0xE4B71251
  <path:IMAGE.md>
  - image name: **R8C5_0x1251_b1**  build id 0xE4B71251  role RING8 C5/ACC22 (lab-host build)  state: WAITING: mesh lab-host-3816811-751322113 (lab-host-main), queued 2026-10-11 16:4x; place feasibility on SLR2 ~97
- image:0xE4B71251_R8C16_0x1251_b1 |  | current | Image 0xE4B71251_R8C16_0x1251_b1
  <path:IMAGE.md>
  - image name: **R8C16_0x1251_b1**  build id 0xE4B71251  role RING8 C16 (lab-host build)  state: FAIL (place DRC UTLZ-1, placer not run): LUT 3,981,216 req vs 2,574,208 device [M]; EXIT rc=1 18:01 EDT; sup
- image:0xE4B71251_R8C5_0x1251_s2 |  | current | Image 0xE4B71251_R8C5_0x1251_s2
  <path:IMAGE.md>
  - image name: **R8C5_0x1251_s2**  build id 0xE4B71251  role RING8 C5/ACC22 seed (HQ)  state: WAITING: mesh build-host-M-753582-751642854 (HQ main 16:4x), dest /mnt/data2/eagle4_scratch/mid/r8_1251_c5_s2  p
- image:0xE4B71251_R8C5_0x1251_s3 |  | current | Image 0xE4B71251_R8C5_0x1251_s3
  <path:IMAGE.md>
  - image name: **R8C5_0x1251_s3**  build id 0xE4B71251  role RING8 C5/ACC22 seed (QM-38)  state: WAITING: mesh build-vm-3652594-753390066 (QM-38 17:16; replaces withdrawn dup s2 build-vm-3651887), dest /
- image:0xE4B71258 |  | current | Image 0xE4B71258
  <path:IMAGE.md>
  - image name: **TUNIT_0x1258_s2**  build id 0xE4B71258  role Track T correctness unit  state: PASS QM-38 (WNS 0.587 / WHS 0.010; fresh PDI, BUILD_ID E4B71258); RELAYED 2026-10-11 19:19 EDT create-only
- image:0xE4B7125D |  | current | Image 0xE4B7125D
  <path:IMAGE.md>
  - image name: **A2A4_LINK**  build id 0xE4B7125D  role LINK (4-card)  state: QM PASS 2026-10-15  pdi md5 de2b963b
- image:0xE4B7125F |  | current | Image 0xE4B7125F
  <path:IMAGE.md>
  - image name: **R8BEST_s1c_s2c**  build id 0xE4B7125F  role HBMS+ring 8-card  state: FAILED 2026-10-15 (impl abort: hbms_cdc.tcl FOLD cells 0)  pdi md5 n/a
- image:0xE4B7125F_R8BEST_kitB |  | current | Image 0xE4B7125F_R8BEST_kitB
  <path:IMAGE.md>
  - image name: **R8BEST_kitB**  build id 0xE4B7125F  role HBMS+ring 8-card (diagnostic)  state: QM FAIL 2026-10-17 (timing)  pdi md5 93b2c619
- image:0xE4B7125F_R8BEST_kitBM |  | current | Image 0xE4B7125F_R8BEST_kitBM
  <path:IMAGE.md>
  - image name: **R8BEST_kitBM**  build id 0xE4B7125F  role HBMS+ring 8-card (diagnostic)  state: QM FAIL 2026-10-17 (timing)  pdi md5 cfed184c
- image:0xE4B7125F_R8BEST_s1c |  | current | Image 0xE4B7125F_R8BEST_s1c
  <path:IMAGE.md>
  - image name: **R8BEST_s1c**  build id 0xE4B7125F  role HBMS+ring 8-card  state: QM FAIL 2026-10-16 (timing)  pdi md5 d9c43417
- image:0xE4B71261 |  | current | Image 0xE4B71261
  <path:IMAGE.md>
  - image name: **R8_AR8_v6**  build id 0xE4B71261  role 8-ring HBM (0x1256 family)  state: BUILDING 2026-10-17  pdi md5 n/a
- image:0xE4B71261_R8_AR8_v6_s1b |  | current | Image 0xE4B71261_R8_AR8_v6_s1b
  <path:IMAGE.md>
  - image name: **R8_AR8_v6_s1b**  build id 0xE4B71261  role 8-ring HBM (0x1256 family)  state: QM FAIL 2026-10-17 (margin)  pdi md5 56593f45
- image:0xE4B71262 |  | current | Image 0xE4B71262
  <path:IMAGE.md>
  - image name: **R8_AR8_v8**  build id 0xE4B71262  role 8-ring HBM (0x1256 family)  state: BUILDING 2026-10-17  pdi md5 n/a
- image:0xE4B71262_R8_AR8_v8_s1 |  | current | Image 0xE4B71262_R8_AR8_v8_s1
  <path:IMAGE.md>
  - image name: **R8_AR8_v8_s1**  build id 0xE4B71262  role 8-ring HBM (0x1256 family) talon 8-card  state: QM PASS + SILICON PASS 2026-10-17 (talon 8/8)  pdi md5 a41a8e90
- image:0xE4B71262_R8_AR8_v8_s1_silicon |  | current | Image 0xE4B71262_R8_AR8_v8_s1_silicon
  <path:IMAGE.md>
  - image name: **R8_AR8_v8_s1_silicon**  build id 0xE4B71262  role 8-ring HBM talon 8-card SILICON  state: SILICON PASS 2026-10-17 [M]  pdi md5 a41a8e90
- image:0xE4B71262_R8_AR8_v8_s1_silicon_p0 |  | current | Image 0xE4B71262_R8_AR8_v8_s1_silicon_p0
  <path:IMAGE.md>
  - image name: **R8_AR8_v8_s1_silicon_p0**  build id 0xE4B71262  role 8-ring HBM talon 8-card SILICON (partition 0)  state: SILICON PASS 2026-10-17 p0 [M]  pdi md5 a41a8e90
- image:0xE4B71262_R8_AR8_v8_s2 |  | current | Image 0xE4B71262_R8_AR8_v8_s2
  <path:IMAGE.md>
  - image name: **R8_AR8_v8_s2**  build id 0xE4B71262  role 8-ring HBM (0x1256 family) talon 8-card  state: QM PASS 2026-10-17  pdi md5 61251d59
- image:0xE4B71262_TALON_LLAMA3B_TP8_v8 |  | current | Image 0xE4B71262_TALON_LLAMA3B_TP8_v8
  <path:IMAGE.md>
  - image name: **TALON_LLAMA3B_TP8_v8**  build id 0xE4B71262  role 8-card talon serving (silicon)  state: SILICON 2026-10-17 [M]  pdi md5 a41a8e90
- image:0xE4B71262_TALON_QWEN27B_W8_TP8_v8 |  | current | Image 0xE4B71262_TALON_QWEN27B_W8_TP8_v8
  <path:IMAGE.md>
  - image name: **TALON_QWEN27B_W8_TP8_v8**  build id 0xE4B71262  role 8-card talon serving (silicon)  state: SILICON 2026-10-18 [M]  pdi md5 a41a8e90
- image:0xE4B71262_TALON_SERVE_v8_suite |  | current | Image 0xE4B71262_TALON_SERVE_v8_suite
  <path:IMAGE.md>
  - image name: **TALON_SERVE_v8_suite**  build id 0xE4B71262  role 8-card talon serving (silicon)  state: SILICON 2026-10-17 [M]  pdi md5 a41a8e90
- image:0xE4B71263 |  | current | Image 0xE4B71263
  <path:IMAGE.md>
  - image name: **R8_AR8_v9b**  build id 0xE4B71263  role 8-ring HBM (0x1256 family)  state: QM PASS 2026-10-17 (thin)  pdi md5 n/a
- image:0xE4B71263_R8_AR8_v9b_s1 |  | current | Image 0xE4B71263_R8_AR8_v9b_s1
  <path:IMAGE.md>
  - image name: **R8_AR8_v9b_s1**  build id 0xE4B71263  role 8-ring HBM (0x1256 family) talon 8-card  state: QM PASS 2026-10-17 (thin)  pdi md5 61fbf1a4
- image:0xE4B71265 A5 INCREMENTAL MARGIN MG_A5IB (K1) |  | current | Image 0xE4B71265 A5 INCREMENTAL MARGIN MG_A5IB (K1)
  <path:0xE4B71265> A5 INCREMENTAL MARGIN MG_A5IB (K1)/IMAGE.md
  - image name: **R8_TARGET_2A_mg_A5Ib**  build id 0xE4B71265 A5 INCREMENTAL MARGIN MG_A5IB (K1)  role 8-card TARGET 2a (track 2 / K1)  state: STOPPED 2026-10-21 by T2 (no PDI; A5 line superseded by 0x1
- image:0xE4B71265 A6 A6R_V1 (K1; A5 + CDC_N + AXM2 2ND SLR1 AXI-SLICE; PLACE EXPLORE; CORE 150) |  | current | Image 0xE4B71265 A6 A6R_V1 (K1; A5 + CDC_N + AXM2 2ND SLR1 AXI-SLICE; PLACE EXPLORE; CORE 150)
  <path:0xE4B71265> A6 A6R_V1 (K1; A5 + CDC_N + AXM2 2ND SLR1 AXI-SLICE; PLACE EXPLORE; CORE 150)/IMAGE.md
  - image name: **R8_TARGET_2A_A6_V1**  build id 0xE4B71265 A6 A6R_V1 (K1; A5 + CDC_N + AXM2 2ND SLR1 AXI-SLICE; PLACE EXPLORE; CORE 150)  role 8-card TARGET 2a (track 2 / K1)  state: QM FAIL 2026-10-20
- image:0xE4B71265 A6 A6R_V3P80 (K1; AE) |  | current | Image 0xE4B71265 A6 A6R_V3P80 (K1; AE)
  <path:0xE4B71265> A6 A6R_V3P80 (K1; AE)/IMAGE.md
  - image name: **R8_TARGET_2A_A6_V3**  build id 0xE4B71265 A6 A6R_V3P80 (K1; AE)  role 8-card TARGET 2a (track 2 / K1)  state: QM FAIL 2026-10-20 (true setup +0.001, hold +0.004)  pdi md5 9a97a10742eeb
- image:0xE4B71265 A6 V1 INCREMENTAL MARGIN + MCP MG_A6V1 (K1) |  | current | Image 0xE4B71265 A6 V1 INCREMENTAL MARGIN + MCP MG_A6V1 (K1)
  <path:0xE4B71265> A6 V1 INCREMENTAL MARGIN + MCP MG_A6V1 (K1)/IMAGE.md
  - image name: **R8_TARGET_2A_mg_A6V1**  build id 0xE4B71265 A6 V1 INCREMENTAL MARGIN + MCP MG_A6V1 (K1)  role 8-card TARGET 2a (track 2 / K1)  state: STOPPED 2026-10-21 08:16Z (VIVADO_EXIT 143, no PDI
- image:0xE4B71265 A6 V3 MARGIN MG_A6V3S (+0.050 MCLK) |  | current | Image 0xE4B71265 A6 V3 MARGIN MG_A6V3S (+0.050 MCLK)
  <path:0xE4B71265> A6 V3 MARGIN MG_A6V3S (+0.050 MCLK)/IMAGE.md
  - image name: **R8_TARGET_2A_mg_A6V3s**  build id 0xE4B71265 A6 V3 MARGIN MG_A6V3S (+0.050 MCLK)  role 8-card TARGET 2a (track 2 / K1)  state: QM FAIL 2026-10-21 (true hbms_mclk +0.027, unchanged from
- image:0xE4B71265 A6 V3 MARGIN RE-ROUTE MG_A6R_V3P80 (K1) |  | current | Image 0xE4B71265 A6 V3 MARGIN RE-ROUTE MG_A6R_V3P80 (K1)
  <path:0xE4B71265> A6 V3 MARGIN RE-ROUTE MG_A6R_V3P80 (K1)/IMAGE.md
  - image name: **R8_TARGET_2A_mg_A6_V3**  build id 0xE4B71265 A6 V3 MARGIN RE-ROUTE MG_A6R_V3P80 (K1)  role 8-card TARGET 2a (track 2 / K1)  state: QM FAIL 2026-10-20 (true hbms_mclk +0.027 < +0.03; WH
- image:0xE4B71265 MG_A5NTR (K1 FULL MARGIN RE-ROUTE OF A5_NTR) |  | current | Image 0xE4B71265 MG_A5NTR (K1 FULL MARGIN RE-ROUTE OF A5_NTR)
  <path:0xE4B71265> MG_A5NTR (K1 FULL MARGIN RE-ROUTE OF A5_NTR)/IMAGE.md
  - image name: **R8_TARGET_2A_mg_A5NTR**  build id 0xE4B71265 MG_A5NTR (K1 FULL MARGIN RE-ROUTE OF A5_NTR)  role 8-card TARGET 2a (track 2 / K1)  state: QM FAIL 2026-10-20 (true -0.040, 1553 failing [M
- image:0xE4B71265 |  | current | Image 0xE4B71265
  <path:IMAGE.md>
  - image name: **R8_DONE_RIGHT_A3**  build id 0xE4B71265  role R8 target 2A Q27B W8 v2 (8-card)  state: QM FAIL 2026-10-19 (timing  pdi md5 +0.009
- image:0xE4B71266 B4 B4Q_V4S (K1; AXM2 B ARM) |  | current | Image 0xE4B71266 B4 B4Q_V4S (K1; AXM2 B ARM)
  <path:0xE4B71266> B4 B4Q_V4S (K1; AXM2 B ARM)/IMAGE.md
  - image name: **R8_TARGET_2A_B4_V4s**  build id 0xE4B71266 B4 B4Q_V4S (K1; AXM2 B ARM)  role 8-card TARGET 2a armB (track 2 / K1)  state: QM FAIL 2026-10-20 (hbms_mclk -0.153 true, 39 failing; WHS 0.0
- image:0xE4B71267 V4 (DS67A5 V4 = A5 TWIN, FCAPS 0X464C4027; NOT DSPARK-SPECIFIC) |  | current | Image 0xE4B71267 V4 (DS67A5 V4 = A5 TWIN, FCAPS 0X464C4027; NOT DSPARK-SPECIFIC)
  <path:0xE4B71267> V4 (DS67A5 V4 = A5 TWIN, FCAPS 0X464C4027; NOT DSPARK-SPECIFIC)/IMAGE.md
  - image name: **DSPARK_2C_A5_v4**  build id 0xE4B71267 V4 (DS67A5 V4 = A5 TWIN, FCAPS 0X464C4027; NOT DSPARK-SPECIFIC)  role 8-card TARGET 2a A5 twin (track 2 / DS)  state: QM FAIL 2026-10-20 (timing:
- image:0xE4B71267 V5 (DSPARK ON A6 BASE, K1 DS5R_V3P80) |  | current | Image 0xE4B71267 V5 (DSPARK ON A6 BASE, K1 DS5R_V3P80)
  <path:0xE4B71267> V5 (DSPARK ON A6 BASE, K1 DS5R_V3P80)/IMAGE.md
  - image name: **DSPARK_v5_ds5r_V3**  build id 0xE4B71267 V5 (DSPARK ON A6 BASE, K1 DS5R_V3P80)  role 8-card TARGET 2a A6 twin / DSpark (track 2)  state: QM FAIL 2026-10-20 (true setup +0.002 < +0.03; 
- image:0xE4B71267 V5 DSPARK (A6 BASE; K1 MARGIN RE-ROUTE MG_DS5R_V3P80 OF DS5R_V3P80 PHYSOPT DCP 82AAE316) |  | current | Image 0xE4B71267 V5 DSPARK (A6 BASE; K1 MARGIN RE-ROUTE MG_DS5R_V3P80 OF DS5R_V3P80 PHYSOPT DCP 82AA
  <path:0xE4B71267> V5 DSPARK (A6 BASE; K1 MARGIN RE-ROUTE MG_DS5R_V3P80 OF DS5R_V3P80 PHYSOPT DCP 82AAE316)/IMAGE.md
  - image name: **DSPARK_v5_mg_ds5r_V3p80**  build id 0xE4B71267 V5 DSPARK (A6 BASE; K1 MARGIN RE-ROUTE MG_DS5R_V3P80 OF DS5R_V3P80 PHYSOPT DCP 82AAE316)  role 8-card TARGET 2a A6-base DSpark (track 2) 
- image:0xE4B71267 |  | current | Image 0xE4B71267
  <path:IMAGE.md>
  - image name: **DSPARK_2C_DRAFTER_v1b**  build id 0xE4B71267  role DSpark 2-card drafter  state: QM FAIL 2026-10-19 (timing  pdi md5 0.000
- image:0xE4B71267_DSPARK_2C_DRAFTER_v2 |  | current | Image 0xE4B71267_DSPARK_2C_DRAFTER_v2
  <path:IMAGE.md>
  - image name: **DSPARK_2C_DRAFTER_v2**  build id 0xE4B71267  role DSpark 2-card drafter  state: QM FAIL 2026-10-19 (timing; 44380 failing)  pdi md5 7a334209
- image:0xE4B71268 H1 (DSPARK_2C_A6D, 1-HEAD SIBLING; DS68H1 B1) |  | current | Image 0xE4B71268 H1 (DSPARK_2C_A6D, 1-HEAD SIBLING; DS68H1 B1)
  <path:0xE4B71268> H1 (DSPARK_2C_A6D, 1-HEAD SIBLING; DS68H1 B1)/IMAGE.md
  - image name: **DSPARK_2C_A6D_H1_0x1268**  build id 0xE4B71268 H1 (DSPARK_2C_A6D, 1-HEAD SIBLING; DS68H1 B1)  role DSpark 2-card A6D (track 2 / DS)  state: QM FAIL 2026-10-20 (true hbms_mclk +0.002, c
- image:0xE4B71268 H1 INCREMENTAL MARGIN + RESET MCP (K1_MG_DS68H1) |  | current | Image 0xE4B71268 H1 INCREMENTAL MARGIN + RESET MCP (K1_MG_DS68H1)
  <path:0xE4B71268> H1 INCREMENTAL MARGIN + RESET MCP (K1_MG_DS68H1)/IMAGE.md
  - image name: **DSPARK_2C_A6D_H1_margin**  build id 0xE4B71268 H1 INCREMENTAL MARGIN + RESET MCP (K1_MG_DS68H1)  role DSpark 2-card A6D (track 2 / DS)  state: QM FAIL 2026-10-21 (true WNS -0.039, 1 fa
- image:0xE4B71268 H2 (DS68H2 B2) |  | current | Image 0xE4B71268 H2 (DS68H2 B2)
  <path:0xE4B71268> H2 (DS68H2 B2)/IMAGE.md
  - image name: **DSPARK_2C_A6D_H2_0x1268**  build id 0xE4B71268 H2 (DS68H2 B2)  role DSpark 2-card A6D (track 2 / DS)  state: QM FAIL 2026-10-21 (true hbms_mclk 0.000; WHS +0.003 design)  pdi md5 2b400
- image:0xE4B7126B B5_MR3 (R1_B5; TREEBLK + MR3) |  | current | Image 0xE4B7126B B5_MR3 (R1_B5; TREEBLK + MR3)
  <path:0xE4B7126B> B5_MR3 (R1_B5; TREEBLK + MR3)/IMAGE.md
  - image name: **R8_B5_MR3_0x126B**  build id 0xE4B7126B B5_MR3 (R1_B5; TREEBLK + MR3)  role 8-card TARGET B5 (track 2 / R1)  state: QM FAIL 2026-10-21 (true hbms_mclk +0.002; WHS +0.007; gate open)  p
- image:0xE4B7126D R2 @225 (K1_R2_225, BUILT WITH MARGIN UU) |  | current | Image 0xE4B7126D R2 @225 (K1_R2_225, BUILT WITH MARGIN UU)
  <path:0xE4B7126D> R2 @225 (K1_R2_225, BUILT WITH MARGIN UU)/IMAGE.md
  - image name: **R8_R2_225_0x126D**  build id 0xE4B7126D R2 @225 (K1_R2_225, BUILT WITH MARGIN UU)  role 8-card TARGET armB R2 (track 2 / K1)  state: QM PASS 2026-10-21 with accepted reset-release MCP 
- image:0xE4B71270 |  | current | Image 0xE4B71270
  <path:IMAGE.md>
  - image name: **4C1_ATT**  build id 0xE4B71270  role 4-card drafter ATT (cards 01+42)  state: BUILDING 2026-10-16 (probe)  pdi md5 n/a
- image:0xE4B71271 |  | current | Image 0xE4B71271
  <path:IMAGE.md>
  - image name: **4C1_MLPV**  build id 0xE4B71271  role 4-card drafter MLPV (cards 61+62)  state: PLANNED 2026-10-16  pdi md5 n/a
- image:0xE4B71271_4C1_ATT_f5 |  | current | Image 0xE4B71271_4C1_ATT_f5
  <path:IMAGE.md>
  - image name: **4C1_ATT_f5**  build id 0xE4B71271  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS only)  pdi md5 0487ad1b
- image:0xE4B71271_4C1_ATT_s3LAattf4 |  | current | Image 0xE4B71271_4C1_ATT_s3LAattf4
  <path:IMAGE.md>
  - image name: **4C1_ATT_s3LAattf4**  build id 0xE4B71271  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-17 (timing)  pdi md5 11544e4f
- image:0xE4B71272 |  | current | Image 0xE4B71272
  <path:IMAGE.md>
  - image name: **4C1_MLPV_s3LCmlpv3**  build id 0xE4B71272  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-17 (timing)  pdi md5 40eeaf3e
- image:0xE4B71272_4C1_MLPV_m4_s3LCmlpvm4f |  | current | Image 0xE4B71272_4C1_MLPV_m4_s3LCmlpvm4f
  <path:IMAGE.md>
  - image name: **4C1_MLPV_m4_s3LCmlpvm4f**  build id 0xE4B71272  role 4-card drafter MLPV m4 (cards 61+62)  state: QM FAIL 2026-10-17 (timing)  pdi md5 72e91468
- image:0xE4B71272_4C1_MLPV_mlpv66 |  | current | Image 0xE4B71272_4C1_MLPV_mlpv66
  <path:IMAGE.md>
  - image name: **4C1_MLPV_mlpv66**  build id 0xE4B71272  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-18 (margin; 4 new FAIL classes to adjudicate)  pdi md5 7987c318
- image:0xE4B71281 |  | current | Image 0xE4B71281
  <path:IMAGE.md>
  - image name: **4C1_ATT_v1**  build id 0xE4B71281  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 a471df87
- image:0xE4B71282 |  | current | Image 0xE4B71282
  <path:IMAGE.md>
  - image name: **4C1_MLPV_x1fix**  build id 0xE4B71282  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (margin)  pdi md5 10a8f70c
- image:0xE4B71282_4C1_MLPV_x1fix_fExp |  | current | Image 0xE4B71282_4C1_MLPV_x1fix_fExp
  <path:IMAGE.md>
  - image name: **4C1_MLPV_x1fix_fExp**  build id 0xE4B71282  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (margin)  pdi md5 7f4ac6c2
- image:0xE4B71282_4C1_MLPV_x1fix_fNtr |  | current | Image 0xE4B71282_4C1_MLPV_x1fix_fNtr
  <path:IMAGE.md>
  - image name: **4C1_MLPV_x1fix_fNtr**  build id 0xE4B71282  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (margin)  pdi md5 7e00c2e4
- image:0xE4B71283 (FANOUT OF ATT_V11) |  | current | Image 0xE4B71283 (FANOUT OF ATT_V11)
  <path:0xE4B71283> (FANOUT OF ATT_V11)/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_explore**  build id 0xE4B71283 (FANOUT OF ATT_V11)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS +0.015 < +0.03)  pdi md5 51bc7811
- image:0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_aexp |  | current | Image 0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_aexp
  <path:0xE4B71283> (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_aexp/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_aexp**  build id 0xE4B71283 (FANOUT OF ATT_V11)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS +0.026 < +0.03)  pdi md5 c12d43c2 (top_
- image:0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_alt |  | current | Image 0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_alt
  <path:0xE4B71283> (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_alt/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_alt**  build id 0xE4B71283 (FANOUT OF ATT_V11)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS +0.021 < +0.03)  pdi md5 53ef7e39 (top_w
- image:0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_eppo |  | current | Image 0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_eppo
  <path:0xE4B71283> (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_eppo/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_eppo**  build id 0xE4B71283 (FANOUT OF ATT_V11)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS +0.021 < +0.03)  pdi md5 36603927 (top_
- image:0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_eto |  | current | Image 0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_eto
  <path:0xE4B71283> (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_eto/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_eto**  build id 0xE4B71283 (FANOUT OF ATT_V11)  role 4-card drafter ATT (cards 01+42)  state: buildx02  pdi md5 +0.010
- image:0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_ntr |  | current | Image 0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_ntr
  <path:0xE4B71283> (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_ntr/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_ntr**  build id 0xE4B71283 (FANOUT OF ATT_V11)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS 0.000 < +0.03)  pdi md5 27cbdeca (top_wr
- image:0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_ssi |  | current | Image 0xE4B71283 (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_ssi
  <path:0xE4B71283> (FANOUT OF ATT_V11)_4C1_ATT_v11_foA11_ssi/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_ssi**  build id 0xE4B71283 (FANOUT OF ATT_V11)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS +0.021 < +0.03)  pdi md5 c6564405 (top_w
- image:0xE4B71283 (FANOUT) |  | current | Image 0xE4B71283 (FANOUT)
  <path:0xE4B71283> (FANOUT)/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_etosub**  build id 0xE4B71283 (FANOUT)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 56b12bf4
- image:0xE4B71283 (FANOUT)_4C1_ATT_v11_foA11_aexpsub |  | current | Image 0xE4B71283 (FANOUT)_4C1_ATT_v11_foA11_aexpsub
  <path:0xE4B71283> (FANOUT)_4C1_ATT_v11_foA11_aexpsub/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_aexpsub**  build id 0xE4B71283 (FANOUT)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 5c4d0d4e
- image:0xE4B71283 (FANOUT)_4C1_ATT_v11_foA11_explore_run2 |  | current | Image 0xE4B71283 (FANOUT)_4C1_ATT_v11_foA11_explore_run2
  <path:0xE4B71283> (FANOUT)_4C1_ATT_v11_foA11_explore_run2/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_explore_run2**  build id 0xE4B71283 (FANOUT)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 652e8124
- image:0xE4B71283 (FANOUT)_4C1_ATT_v11_foA11_expsub |  | current | Image 0xE4B71283 (FANOUT)_4C1_ATT_v11_foA11_expsub
  <path:0xE4B71283> (FANOUT)_4C1_ATT_v11_foA11_expsub/IMAGE.md
  - image name: **4C1_ATT_v11_foA11_expsub**  build id 0xE4B71283 (FANOUT)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (timing  pdi md5 0.000
- image:0xE4B71283 |  | current | Image 0xE4B71283
  <path:IMAGE.md>
  - image name: **4C1_ATT_v11_s3LAattv11**  build id 0xE4B71283  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-18 (margin: WNS 0.000 < +0.03)  pdi md5 c8463ff3
- image:0xE4B71284 (FANOUT) |  | current | Image 0xE4B71284 (FANOUT)
  <path:0xE4B71284> (FANOUT)/IMAGE.md
  - image name: **4C1_MLPV_0x1284_m12c_ntr_hf**  build id 0xE4B71284 (FANOUT)  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (margin)  pdi md5 71dc9537
- image:0xE4B71284 (FANOUT)_4C1_MLPV_0x1284_m12c_ae_hf |  | current | Image 0xE4B71284 (FANOUT)_4C1_MLPV_0x1284_m12c_ae_hf
  <path:0xE4B71284> (FANOUT)_4C1_MLPV_0x1284_m12c_ae_hf/IMAGE.md
  - image name: **4C1_MLPV_0x1284_m12c_ae_hf**  build id 0xE4B71284 (FANOUT)  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (margin)  pdi md5 b5ecbd16
- image:0xE4B71284 MLPV V12 (S3LCMLPV12C) |  | current | Image 0xE4B71284 MLPV V12 (S3LCMLPV12C)
  <path:0xE4B71284> MLPV V12 (S3LCMLPV12C)/IMAGE.md
  - image name: **4C1_MLPV_v12c**  build id 0xE4B71284 MLPV V12 (S3LCMLPV12C)  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (margin)  pdi md5 0eca56c1
- image:0xE4B71287 |  | current | Image 0xE4B71287
  <path:IMAGE.md>
  - image name: **4C1_MLPV_allfix_D3**  build id 0xE4B71287  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (margin)  pdi md5 b2333158
- image:0xE4B71288 (THIN-MARGIN RE-ROUTE AE) |  | current | Image 0xE4B71288 (THIN-MARGIN RE-ROUTE AE)
  <path:0xE4B71288> (THIN-MARGIN RE-ROUTE AE)/IMAGE.md
  - image name: **4C1_ATT_0x1288_foTH_A88ae**  build id 0xE4B71288 (THIN-MARGIN RE-ROUTE AE)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-20 (setup margin: +0.006)  pdi md5 b1fac18e
- image:0xE4B71288 |  | current | Image 0xE4B71288
  <path:IMAGE.md>
  - image name: **4C1_ATT_v11b2**  build id 0xE4B71288  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 1b449de3
- image:0xE4B71289 (THIN-MARGIN RE-PLACE) |  | current | Image 0xE4B71289 (THIN-MARGIN RE-PLACE)
  <path:0xE4B71289> (THIN-MARGIN RE-PLACE)/IMAGE.md
  - image name: **4C1_ATT_0x1289_foTH_A89mg_kvf**  build id 0xE4B71289 (THIN-MARGIN RE-PLACE)  role 4-card drafter ATT (cards 01+42)  state: QM PASS 2026-10-19 (timing: true WNS +0.065 / WHS +0.011; FAI
- image:0xE4B71289 |  | current | Image 0xE4B71289
  <path:IMAGE.md>
  - image name: **4C1_ATT_v11d2**  build id 0xE4B71289  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (timing  pdi md5 +0.006
- image:0xE4B7128A |  | current | Image 0xE4B7128A
  <path:IMAGE.md>
  - image name: **4C1_ATT_v11p2**  build id 0xE4B7128A  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 d1072545
- image:0xE4B7128B (THIN-MARGIN RE-PLACE) |  | current | Image 0xE4B7128B (THIN-MARGIN RE-PLACE)
  <path:0xE4B7128B> (THIN-MARGIN RE-PLACE)/IMAGE.md
  - image name: **4C1_MLPV_v12frst_foTH_M8B_rp**  build id 0xE4B7128B (THIN-MARGIN RE-PLACE)  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-20 (margin: +0.021)  pdi md5 bf0d44b8
- image:0xE4B7128B (THIN-MARGIN RE-ROUTE) |  | current | Image 0xE4B7128B (THIN-MARGIN RE-ROUTE)
  <path:0xE4B7128B> (THIN-MARGIN RE-ROUTE)/IMAGE.md
  - image name: **4C1_MLPV_v12frst_foTH_M8B_mg**  build id 0xE4B7128B (THIN-MARGIN RE-ROUTE)  role 4-card drafter MLPV (cards 61+62)  state: QM PASS 2026-10-19 (timing: true WNS +0.030 / WHS +0.010; FAI
- image:0xE4B7128C |  | current | Image 0xE4B7128C
  <path:IMAGE.md>
  - image name: **4C1_ATT_v11pi**  build id 0xE4B7128C  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 6880d151
- image:0xE4B7128F |  | current | Image 0xE4B7128F
  <path:IMAGE.md>
  - image name: **4C1_ATT_v11pX**  build id 0xE4B7128F  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (margin)  pdi md5 ba6b24a5
- image:0xE4B71291 (lab-host TP4 DEADHOP-HOLD RE-ROUTE FOTH_T4_DH) |  | current | Image 0xE4B71291 (lab-host TP4 DEADHOP-HOLD RE-ROUTE FOTH_T4_DH)
  <path:0xE4B71291> (lab-host TP4 DEADHOP-HOLD RE-ROUTE FOTH_T4_DH)/IMAGE.md
  - image name: **4C1_TP4_foTH_T4_dh**  build id 0xE4B71291 (lab-host TP4 DEADHOP-HOLD RE-ROUTE FOTH_T4_DH)  role 4-card TP4 link (lab-host)  state: QM FAIL 2026-10-20 (usr_0 -0.264, 28203 failing; WHS 0.000)  
- image:0xE4B71291 (lab-host TP4V01C FOUR-CARD TP4 LINK BD; MESH CLASS PDI_BUILD_FOURCARD_TP4) |  | current | Image 0xE4B71291 (lab-host TP4V01C FOUR-CARD TP4 LINK BD; MESH CLASS PDI_BUILD_FOURCARD_TP4)
  <path:0xE4B71291> (lab-host TP4V01C FOUR-CARD TP4 LINK BD; MESH CLASS PDI_BUILD_FOURCARD_TP4)/IMAGE.md
  - image name: **4C1_TP4_v01c**  build id 0xE4B71291 (lab-host TP4V01C FOUR-CARD TP4 LINK BD; MESH CLASS PDI_BUILD_FOURCARD_TP4)  role 4-card TP4 link (lab-host)  state: QM FAIL 2026-10-20 (timing: WNS -9.062,
- image:0xE4B71292 (lab-host TP4V01D) |  | current | Image 0xE4B71292 (lab-host TP4V01D)
  <path:0xE4B71292> (lab-host TP4V01D)/IMAGE.md
  - image name: **4C1_TP4_v01d**  build id 0xE4B71292 (lab-host TP4V01D)  role 4-card TP4  state: QM FAIL 2026-10-20 (timing: 120181 failing; hold 14 failing)  pdi md5 17a03248
- image:0xE4B7129A (lab-host S3LADUALQ1B; UNI_ONEFLIP COPY 0XE4B7129A_66M_9B23B3D6) |  | current | Image 0xE4B7129A (lab-host S3LADUALQ1B; UNI_ONEFLIP COPY 0XE4B7129A_66M_9B23B3D6)
  <path:0xE4B7129A> (lab-host S3LADUALQ1B; UNI_ONEFLIP COPY 0XE4B7129A_66M_9B23B3D6)/IMAGE.md
  - image name: **4C1_ATT_dualq1b**  build id 0xE4B7129A (lab-host S3LADUALQ1B; UNI_ONEFLIP COPY 0XE4B7129A_66M_9B23B3D6)  role 4-card drafter ATT  state: QM FAIL 2026-10-21 (0.000 / +0.010)  pdi md5 12a981
- image:0xE4B7129A (MARGIN RE-ROUTE OF lab-host S3LADUALQ1B PHYSOPT DCP DD88AF35) |  | current | Image 0xE4B7129A (MARGIN RE-ROUTE OF lab-host S3LADUALQ1B PHYSOPT DCP DD88AF35)
  <path:0xE4B7129A> (MARGIN RE-ROUTE OF lab-host S3LADUALQ1B PHYSOPT DCP DD88AF35)/IMAGE.md
  - image name: **4C1_ATT_dualq1b_foTH_A9A_mg**  build id 0xE4B7129A (MARGIN RE-ROUTE OF lab-host S3LADUALQ1B PHYSOPT DCP DD88AF35)  role 4-card drafter ATT (cards 01+42)  state: QM PASS 2026-10-21 (true +0
- image:0xE4B712A1 (lab-host S3LACTRLA; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712A1_66M_82F0124F.PDI) |  | current | Image 0xE4B712A1 (lab-host S3LACTRLA; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712A1_66M_82F0124F.P
  <path:0xE4B712A1> (lab-host S3LACTRLA; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712A1_66M_82F0124F.PDI)/IMAGE.md
  - image name: **4C1_ATT_ctrlA**  build id 0xE4B712A1 (lab-host S3LACTRLA; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712A1_66M_82F0124F.PDI)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-
- image:0xE4B712A2 (MARGIN RE-ROUTE OF lab-host S3LCCTRLA; ID ALSO USED BY HQ URT_ACB16 IN RING8_CONTROL ROW 12 - IDENTIFY BY MD5) |  | current | Image 0xE4B712A2 (MARGIN RE-ROUTE OF lab-host S3LCCTRLA; ID ALSO USED BY HQ URT_ACB16 IN RING8_CONTROL R
  <path:0xE4B712A2> (MARGIN RE-ROUTE OF lab-host S3LCCTRLA; ID ALSO USED BY HQ URT_ACB16 IN RING8_CONTROL ROW 12 - IDENTIFY BY MD5)/IMAGE.md
  - image name: **4C1_MLPV_ctrlA_foTH_A2_mg**  build id 0xE4B712A2 (MARGIN RE-ROUTE OF lab-host S3LCCTRLA; ID ALSO USED BY HQ URT_ACB16 IN RING8_CONTROL ROW 12 - IDENTIFY BY MD5)  role 4-card drafter MLPV (
- image:0xE4B712B1 (lab-host S3LAKVCA ATT HEAD; FINAL PDI 5D35CB3B) |  | current | Image 0xE4B712B1 (lab-host S3LAKVCA ATT HEAD; FINAL PDI 5D35CB3B)
  <path:0xE4B712B1> (lab-host S3LAKVCA ATT HEAD; FINAL PDI 5D35CB3B)/IMAGE.md
  - image name: **4C1_ATT_kvcA**  build id 0xE4B712B1 (lab-host S3LAKVCA ATT HEAD; FINAL PDI 5D35CB3B)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-19 (setup + hold margin)  pdi md5 f23a6b
- image:0xE4B712B1 (MARGIN RE-ROUTE OF lab-host S3LAKVCA) |  | current | Image 0xE4B712B1 (MARGIN RE-ROUTE OF lab-host S3LAKVCA)
  <path:0xE4B712B1> (MARGIN RE-ROUTE OF lab-host S3LAKVCA)/IMAGE.md
  - image name: **4C1_ATT_kvcA_foTH_B1_mg**  build id 0xE4B712B1 (MARGIN RE-ROUTE OF lab-host S3LAKVCA)  role 4-card drafter ATT (cards 01+42)  state: QM PASS 2026-10-20 (true WNS +0.044 / WHS +0.011; 0 new
- image:0xE4B712B2 (lab-host KVF BUILD; UNI_ONEFLIP COPY ON build-host-M) |  | current | Image 0xE4B712B2 (lab-host KVF BUILD; UNI_ONEFLIP COPY ON build-host-M)
  <path:0xE4B712B2> (lab-host KVF BUILD; UNI_ONEFLIP COPY ON build-host-M)/IMAGE.md
  - image name: **4C1_unknown_0x12B2_build-host-M**  build id 0xE4B712B2 (lab-host KVF BUILD; UNI_ONEFLIP COPY ON build-host-M)  role 4C1 (lab-host)  state: QM FAIL 2026-10-20 (setup +0.005 [M per lab-host TO_HQ 19:21]; sour
- image:0xE4B712B2 (lab-host S3LAKVC2B; FPT F7A78970) |  | current | Image 0xE4B712B2 (lab-host S3LAKVC2B; FPT F7A78970)
  <path:0xE4B712B2> (lab-host S3LAKVC2B; FPT F7A78970)/IMAGE.md
  - image name: **4C1_ATT_kvc2b**  build id 0xE4B712B2 (lab-host S3LAKVC2B; FPT F7A78970)  role 4-card drafter ATT  state: QM FAIL 2026-10-21 (usr_1 0.000 / hold +0.001)  pdi md5 fpt f7a78970
- image:0xE4B712B2 (lab-host S3LAKVC2E; FPT CF5B02FA) |  | current | Image 0xE4B712B2 (lab-host S3LAKVC2E; FPT CF5B02FA)
  <path:0xE4B712B2> (lab-host S3LAKVC2E; FPT CF5B02FA)/IMAGE.md
  - image name: **4C1_ATT_kvc2e**  build id 0xE4B712B2 (lab-host S3LAKVC2E; FPT CF5B02FA)  role 4-card drafter ATT  state: QM FAIL 2026-10-21 (usr_2 +0.004)  pdi md5 fpt cf5b02fa
- image:0xE4B712B3 (lab-host S3LAR3EB3; UNI_ONEFLIP COPY 0XE4B712B3_66M_E41D72CE) |  | current | Image 0xE4B712B3 (lab-host S3LAR3EB3; UNI_ONEFLIP COPY 0XE4B712B3_66M_E41D72CE)
  <path:0xE4B712B3> (lab-host S3LAR3EB3; UNI_ONEFLIP COPY 0XE4B712B3_66M_E41D72CE)/IMAGE.md
  - image name: **4C1_ATT_r3eB3**  build id 0xE4B712B3 (lab-host S3LAR3EB3; UNI_ONEFLIP COPY 0XE4B712B3_66M_E41D72CE)  role 4-card drafter ATT  state: QM FAIL 2026-10-21 (0.000 / +0.010)  pdi md5 12c23c8631
- image:0xE4B712B4 (lab-host KV COMMIT + R3_EARLY + NORMFUSE; UNI_ONEFLIP COPY ON build-host-M) |  | current | Image 0xE4B712B4 (lab-host KV COMMIT + R3_EARLY + NORMFUSE; UNI_ONEFLIP COPY ON build-host-M)
  <path:0xE4B712B4> (lab-host KV COMMIT + R3_EARLY + NORMFUSE; UNI_ONEFLIP COPY ON build-host-M)/IMAGE.md
  - image name: **4C1_unknown_0x12B4_build-host-M**  build id 0xE4B712B4 (lab-host KV COMMIT + R3_EARLY + NORMFUSE; UNI_ONEFLIP COPY ON build-host-M)  role 4C1 (lab-host)  state: QM FAIL 2026-10-20 (setup +0.013 [M per b
- image:0xE4B712B4 (RE-PLACE + MARGIN OF lab-host 0X12B4 OPT DCP 6EB65E80 + KVF SOFT PBLOCK; KV COMMIT + R3_EARLY + NORMFUSE) |  | current | Image 0xE4B712B4 (RE-PLACE + MARGIN OF lab-host 0X12B4 OPT DCP 6EB65E80 + KVF SOFT PBLOCK; KV COMMIT + R
  <path:0xE4B712B4> (RE-PLACE + MARGIN OF lab-host 0X12B4 OPT DCP 6EB65E80 + KVF SOFT PBLOCK; KV COMMIT + R3_EARLY + NORMFUSE)/IMAGE.md
  - image name: **4C1_ATT_B4_foTH_B4_kvf**  build id 0xE4B712B4 (RE-PLACE + MARGIN OF lab-host 0X12B4 OPT DCP 6EB65E80 + KVF SOFT PBLOCK; KV COMMIT + R3_EARLY + NORMFUSE)  role 4-card drafter ATT (cards 01+
- image:0xE4B712BB (lab-host S3LAGKB = G5KVC; UNI_ONEFLIP COPY 0XE4B712BB_66M_6943D8A4) |  | current | Image 0xE4B712BB (lab-host S3LAGKB = G5KVC; UNI_ONEFLIP COPY 0XE4B712BB_66M_6943D8A4)
  <path:0xE4B712BB> (lab-host S3LAGKB = G5KVC; UNI_ONEFLIP COPY 0XE4B712BB_66M_6943D8A4)/IMAGE.md
  - image name: **4C1_ATT_gkB**  build id 0xE4B712BB (lab-host S3LAGKB = G5KVC; UNI_ONEFLIP COPY 0XE4B712BB_66M_6943D8A4)  role 4-card drafter ATT  state: QM FAIL 2026-10-21 (+0.003 / +0.010)  pdi md5 63ab6
- image:0xE4B712C1 (lab-host S3LCMLPV_CONSA; ON SHARED HOST BUILDXD00) |  | current | Image 0xE4B712C1 (lab-host S3LCMLPV_CONSA; ON SHARED HOST BUILDXD00)
  <path:0xE4B712C1> (lab-host S3LCMLPV_CONSA; ON SHARED HOST BUILDXD00)/IMAGE.md
  - image name: **4C1_MLPV_consA**  build id 0xE4B712C1 (lab-host S3LCMLPV_CONSA; ON SHARED HOST BUILDXD00)  role 4C1 MLPV (cards 61+62, lab-host)  state: QM FAIL 2026-10-20 (setup 0.000 and hold +0.008 < +0.01
- image:0xE4B712C2 (lab-host S3LCMLPV_CONSB; ID PER lab-host TO_HQ, ON T3 HOST BUILDXD01) |  | current | Image 0xE4B712C2 (lab-host S3LCMLPV_CONSB; ID PER lab-host TO_HQ, ON T3 HOST BUILDXD01)
  <path:0xE4B712C2> (lab-host S3LCMLPV_CONSB; ID PER lab-host TO_HQ, ON T3 HOST BUILDXD01)/IMAGE.md
  - image name: **4C1_MLPV_consB**  build id 0xE4B712C2 (lab-host S3LCMLPV_CONSB; ID PER lab-host TO_HQ, ON T3 HOST BUILDXD01)  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (setup + hold ma
- image:0xE4B712C3 (lab-host S3LCMLPV_CONSC; ON T3 HOST BUILDXD02) |  | current | Image 0xE4B712C3 (lab-host S3LCMLPV_CONSC; ON T3 HOST BUILDXD02)
  <path:0xE4B712C3> (lab-host S3LCMLPV_CONSC; ON T3 HOST BUILDXD02)/IMAGE.md
  - image name: **4C1_MLPV_consC**  build id 0xE4B712C3 (lab-host S3LCMLPV_CONSC; ON T3 HOST BUILDXD02)  role 4-card drafter MLPV (cards 61+62)  state: QM FAIL 2026-10-19 (setup margin)  pdi md5 18be69e5
- image:0xE4B712C3 (MARGIN RE-ROUTE FOTH_C3_MG) |  | current | Image 0xE4B712C3 (MARGIN RE-ROUTE FOTH_C3_MG)
  <path:0xE4B712C3> (MARGIN RE-ROUTE FOTH_C3_MG)/IMAGE.md
  - image name: **4C1_MLPV_foTH_C3_mg**  build id 0xE4B712C3 (MARGIN RE-ROUTE FOTH_C3_MG)  role 4C1 MLPV (lab-host)  state: QM FAIL 2026-10-20 (setup margin: true +0.001 < +0.03)  pdi md5 c8df752a22ccbfe6d5
- image:0xE4B712C4 (lab-host S3LAATT_CONSA; ON SHARED HOST BUILDXD03) |  | current | Image 0xE4B712C4 (lab-host S3LAATT_CONSA; ON SHARED HOST BUILDXD03)
  <path:0xE4B712C4> (lab-host S3LAATT_CONSA; ON SHARED HOST BUILDXD03)/IMAGE.md
  - image name: **4C1_ATT_consA**  build id 0xE4B712C4 (lab-host S3LAATT_CONSA; ON SHARED HOST BUILDXD03)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-20 (setup margin +0.001 and hold +0.0
- image:0xE4B712C5 (lab-host S3LAATT_CONSB; ON SHARED HOST BUILDXD04) |  | current | Image 0xE4B712C5 (lab-host S3LAATT_CONSB; ON SHARED HOST BUILDXD04)
  <path:0xE4B712C5> (lab-host S3LAATT_CONSB; ON SHARED HOST BUILDXD04)/IMAGE.md
  - image name: **4C1_ATT_consB**  build id 0xE4B712C5 (lab-host S3LAATT_CONSB; ON SHARED HOST BUILDXD04)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-20 (timing: 28660 failing)  pdi md5 0
- image:0xE4B712C6 (lab-host S3LACC6B; UNI_ONEFLIP COPY 0XE4B712C6_66M_1C14E117) |  | current | Image 0xE4B712C6 (lab-host S3LACC6B; UNI_ONEFLIP COPY 0XE4B712C6_66M_1C14E117)
  <path:0xE4B712C6> (lab-host S3LACC6B; UNI_ONEFLIP COPY 0XE4B712C6_66M_1C14E117)/IMAGE.md
  - image name: **4C1_ATT_cc6b**  build id 0xE4B712C6 (lab-host S3LACC6B; UNI_ONEFLIP COPY 0XE4B712C6_66M_1C14E117)  role 4-card drafter ATT  state: QM FAIL 2026-10-21 (usr_2 +0.004)  pdi md5 fpt 1c14e117
- image:0xE4B712C6 (lab-host UNI_ONEFLIP COPY ON build-host-M) |  | current | Image 0xE4B712C6 (lab-host UNI_ONEFLIP COPY ON build-host-M)
  <path:0xE4B712C6> (lab-host UNI_ONEFLIP COPY ON build-host-M)/IMAGE.md
  - image name: **4C1_unknown_0x12C6_build-host-M**  build id 0xE4B712C6 (lab-host UNI_ONEFLIP COPY ON build-host-M)  role 4C1 (lab-host)  state: QM PENDING (source run not seen by HQ)  pdi md5 3047ff86
- image:0xE4B712D0 (lab-host S3LAATT_CONSKV; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712D0_66M_1190F7DF.PDI) |  | current | Image 0xE4B712D0 (lab-host S3LAATT_CONSKV; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712D0_66M_1190F
  <path:0xE4B712D0> (lab-host S3LAATT_CONSKV; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712D0_66M_1190F7DF.PDI)/IMAGE.md
  - image name: **4C1_ATT_consKV**  build id 0xE4B712D0 (lab-host S3LAATT_CONSKV; UNI_ONEFLIP COPY ONEFLIP_P2S_HEAD_build-host-A_0XE4B712D0_66M_1190F7DF.PDI)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL
- image:0xE4B712D1 (2ND MARGIN RE-ROUTE OF lab-host S3LACONS_FULL PHYSOPT DCP 643B90BA; MARGIN_ATT_D1 INCL CLK_PL_2) |  | current | Image 0xE4B712D1 (2ND MARGIN RE-ROUTE OF lab-host S3LACONS_FULL PHYSOPT DCP 643B90BA; MARGIN_ATT_D1 INCL
  <path:0xE4B712D1> (2ND MARGIN RE-ROUTE OF lab-host S3LACONS_FULL PHYSOPT DCP 643B90BA; MARGIN_ATT_D1 INCL CLK_PL_2)/IMAGE.md
  - image name: **4C1_ATT_cons_full_foTH_D1_mg2**  build id 0xE4B712D1 (2ND MARGIN RE-ROUTE OF lab-host S3LACONS_FULL PHYSOPT DCP 643B90BA; MARGIN_ATT_D1 INCL CLK_PL_2)  role 4-card drafter ATT (cards 01+42
- image:0xE4B712D1 (lab-host S3LACONS_FULL; ON SHARED HOST BUILDZ03) |  | current | Image 0xE4B712D1 (lab-host S3LACONS_FULL; ON SHARED HOST BUILDZ03)
  <path:0xE4B712D1> (lab-host S3LACONS_FULL; ON SHARED HOST BUILDZ03)/IMAGE.md
  - image name: **4C1_ATT_cons_full**  build id 0xE4B712D1 (lab-host S3LACONS_FULL; ON SHARED HOST BUILDZ03)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-20 (setup margin: 0.000)  pdi md5 
- image:0xE4B712D1 (MARGIN RE-ROUTE OF lab-host S3LACONS_FULL) |  | current | Image 0xE4B712D1 (MARGIN RE-ROUTE OF lab-host S3LACONS_FULL)
  <path:0xE4B712D1> (MARGIN RE-ROUTE OF lab-host S3LACONS_FULL)/IMAGE.md
  - image name: **4C1_ATT_cons_full_foTH_D1_mg**  build id 0xE4B712D1 (MARGIN RE-ROUTE OF lab-host S3LACONS_FULL)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-20 (setup +0.006 < +0.03; bin
- image:0xE4B712E0 |  | current | Image 0xE4B712E0
  <path:IMAGE.md>
  - image name: **URE1**  build id 0xE4B712E0  role URE HBM streaming  state: QM FAIL 2026-10-16 (timing)  pdi md5 e03ceecd
- image:0xE4B712E0_URE_v3_C2 |  | current | Image 0xE4B712E0_URE_v3_C2
  <path:IMAGE.md>
  - image name: **URE_v3_C2**  build id 0xE4B712E0  role URE v3 C2 (225 MHz)  state: QM FAIL 2026-10-19 (timing  pdi md5 +0.002
- image:0xE4B712E1 (lab-host S3LCG11F BASE; FPT CEC3A8E1) |  | current | Image 0xE4B712E1 (lab-host S3LCG11F BASE; FPT CEC3A8E1)
  <path:0xE4B712E1> (lab-host S3LCG11F BASE; FPT CEC3A8E1)/IMAGE.md
  - image name: **4C1_MLPV_g11f**  build id 0xE4B712E1 (lab-host S3LCG11F BASE; FPT CEC3A8E1)  role 4-card drafter MLPV  state: QM FAIL 2026-10-21 (rxoutclk_3 0.000, usr_1 hold +0.002)  pdi md5 fpt cec3a8e1
- image:0xE4B712E1 (lab-host S3LCG11FR RE-RUN; UNI_ONEFLIP COPY 0XE4B712E1_66M_D13A8F19) |  | current | Image 0xE4B712E1 (lab-host S3LCG11FR RE-RUN; UNI_ONEFLIP COPY 0XE4B712E1_66M_D13A8F19)
  <path:0xE4B712E1> (lab-host S3LCG11FR RE-RUN; UNI_ONEFLIP COPY 0XE4B712E1_66M_D13A8F19)/IMAGE.md
  - image name: **4C1_MLPV_g11fR**  build id 0xE4B712E1 (lab-host S3LCG11FR RE-RUN; UNI_ONEFLIP COPY 0XE4B712E1_66M_D13A8F19)  role 4-card drafter MLPV  state: QM FAIL 2026-10-21 (0.000 / +0.002)  pdi md5 0
- image:0xE4B712E2 (lab-host S3LAG5FA BASE; UNI_ONEFLIP COPY 0XE4B712E2_66M_427D080E) |  | current | Image 0xE4B712E2 (lab-host S3LAG5FA BASE; UNI_ONEFLIP COPY 0XE4B712E2_66M_427D080E)
  <path:0xE4B712E2> (lab-host S3LAG5FA BASE; UNI_ONEFLIP COPY 0XE4B712E2_66M_427D080E)/IMAGE.md
  - image name: **4C1_ATT_g5fA**  build id 0xE4B712E2 (lab-host S3LAG5FA BASE; UNI_ONEFLIP COPY 0XE4B712E2_66M_427D080E)  role 4-card drafter ATT  state: QM FAIL 2026-10-21 (-0.355 / 0.000)  pdi md5 726db1a
- image:0xE4B712E2 (lab-host S3LAG5FC BASE; IDENTIFY BY MD5) |  | current | Image 0xE4B712E2 (lab-host S3LAG5FC BASE; IDENTIFY BY MD5)
  <path:0xE4B712E2> (lab-host S3LAG5FC BASE; IDENTIFY BY MD5)/IMAGE.md
  - image name: **4C1_ATT_g5fC**  build id 0xE4B712E2 (lab-host S3LAG5FC BASE; IDENTIFY BY MD5)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-20 (timing -0.705; WHS 0.000)  pdi md5 5fb76613
- image:0xE4B712E2 (lab-host S3LAG5FD BASE IMPL_1; UNI_ONEFLIP COPY 0XE4B712E2_66M_2434F1B1) |  | current | Image 0xE4B712E2 (lab-host S3LAG5FD BASE IMPL_1; UNI_ONEFLIP COPY 0XE4B712E2_66M_2434F1B1)
  <path:0xE4B712E2> (lab-host S3LAG5FD BASE IMPL_1; UNI_ONEFLIP COPY 0XE4B712E2_66M_2434F1B1)/IMAGE.md
  - image name: **4C1_ATT_g5fD**  build id 0xE4B712E2 (lab-host S3LAG5FD BASE IMPL_1; UNI_ONEFLIP COPY 0XE4B712E2_66M_2434F1B1)  role 4-card drafter ATT (cards 01+42)  state: QM FAIL 2026-10-20 (timing -0.1
- image:0xE4B712E2 (FOTH_E2KVFAE REPEAT ON BUILDX02) |  | current | Image 0xE4B712E2 (FOTH_E2KVFAE REPEAT ON BUILDX02)
  <path:0xE4B712E2> (FOTH_E2KVFAE REPEAT ON BUILDX02)/IMAGE.md
  - image name: **4C1_ATT_E2_foTH_E2kvfae_x02**  build id 0xE4B712E2 (FOTH_E2KVFAE REPEAT ON BUILDX02)  role 4-card drafter ATT (cards 01+42)  state: QM PASS 2026-10-20 (true +0.040 / +0.011; twin of d1
- image:0xE4B712E2 (RE-PLACE + MARGIN OF lab-host 0X12E2 OPT DCP 3EFDE399 WITH KVF_CH0_SLR0 SOFT PBLOCK) |  | current | Image 0xE4B712E2 (RE-PLACE + MARGIN OF lab-host 0X12E2 OPT DCP 3EFDE399 WITH KVF_CH0_SLR0 SOFT PBLOCK)
  <path:0xE4B712E2> (RE-PLACE + MARGIN OF lab-host 0X12E2 OPT DCP 3EFDE399 WITH KVF_CH0_SLR0 SOFT PBLOCK)/IMAGE.md
  - image name: **4C1_ATT_E2_foTH_E2kvf**  build id 0xE4B712E2 (RE-PLACE + MARGIN OF lab-host 0X12E2 OPT DCP 3EFDE399 WITH KVF_CH0_SLR0 SOFT PBLOCK)  role 4-card drafter ATT (cards 01+42)  state: QM PASS 20
- image:0xE4B712E2 (SIBLING OF FOTH_E2KVF, AE PLACE VARIANT) |  | current | Image 0xE4B712E2 (SIBLING OF FOTH_E2KVF, AE PLACE VARIANT)
  <path:0xE4B712E2> (SIBLING OF FOTH_E2KVF, AE PLACE VARIANT)/IMAGE.md
  - image name: **4C1_ATT_E2_foTH_E2kvfae**  build id 0xE4B712E2 (SIBLING OF FOTH_E2KVF, AE PLACE VARIANT)  role 4-card drafter ATT (cards 01+42)  state: QM PASS 2026-10-20 (true WNS +0.040 / WHS +0.011
- image:0xE4B712E3 (lab-host FOTH_E3_RL_ BALSLR RE-PLACE + LMDOWN_RELOC_SLR12 FROM OPT DCP 302DBF35 + MARGIN_MLPV) |  | current | Image 0xE4B712E3 (lab-host FOTH_E3_RL_ BALSLR RE-PLACE + LMDOWN_RELOC_SLR12 FROM OPT DCP 302DBF35 + MARG
  <path:0xE4B712E3> (lab-host FOTH_E3_RL_ BALSLR RE-PLACE + LMDOWN_RELOC_SLR12 FROM OPT DCP 302DBF35 + MARGIN_MLPV)/IMAGE.md
  - image name: **4C1_MLPV_E3_foTH_E3_rl**  build id 0xE4B712E3 (lab-host FOTH_E3_RL: BALSLR RE-PLACE + LMDOWN_RELOC_SLR12 FROM OPT DCP 302DBF35 + MARGIN_MLPV)  role 4-card drafter MLPV  state: QM FAIL 2026
- image:0xE4B712E4 (lab-host S3LCG11E4 BASE; FPT 4A185FDA) |  | current | Image 0xE4B712E4 (lab-host S3LCG11E4 BASE; FPT 4A185FDA)
  <path:0xE4B712E4> (lab-host S3LCG11E4 BASE; FPT 4A185FDA)/IMAGE.md
  - image name: **4C1_MLPV_g11e4**  build id 0xE4B712E4 (lab-host S3LCG11E4 BASE; FPT 4A185FDA)  role 4-card drafter MLPV  state: QM FAIL 2026-10-21 (usr_1 +0.003)  pdi md5 fpt 4a185fda
- image:0xE4B712E4 (ON-CARD 0XE4B7116A) V2 |  | current | Image 0xE4B712E4 (ON-CARD 0XE4B7116A) V2
  <path:0xE4B712E4> (ON-CARD 0XE4B7116A) V2/IMAGE.md
  - image name: **P2B_HEAD_V2**  build id 0xE4B712E4 (ON-CARD 0XE4B7116A) V2  role 2b HEAD  state: QM FAIL 2026-10-19 (setup margin: true WNS +0.002)  pdi md5 9487a151
- image:0xE4B712E4 (ON-CARD 0XE4B7116A) |  | current | Image 0xE4B712E4 (ON-CARD 0XE4B7116A)
  <path:0xE4B712E4> (ON-CARD 0XE4B7116A)/IMAGE.md
  - image name: **P2B_HEAD_impl1**  build id 0xE4B712E4 (ON-CARD 0XE4B7116A)  role 2b HEAD  state: QM FAIL 2026-10-19 (margin)  pdi md5 341db1fd
- image:0xE4B712E4 M3 (MARGIN RE-ROUTE OF V5; ON-CARD 0XE4B7116A) |  | current | Image 0xE4B712E4 M3 (MARGIN RE-ROUTE OF V5; ON-CARD 0XE4B7116A)
  <path:0xE4B712E4> M3 (MARGIN RE-ROUTE OF V5; ON-CARD 0XE4B7116A)/IMAGE.md
  - image name: **P2B_HEAD_M3**  build id 0xE4B712E4 M3 (MARGIN RE-ROUTE OF V5; ON-CARD 0XE4B7116A)  role 2b HEAD  state: QM PASS 2026-10-20 (timing: true WNS +0.044 / WHS +0.021; FAIL classes = 0x12E4 
- image:0xE4B712E4 V2H (HOOKED WDI REPLAY) |  | current | Image 0xE4B712E4 V2H (HOOKED WDI REPLAY)
  <path:0xE4B712E4> V2H (HOOKED WDI REPLAY)/IMAGE.md
  - image name: **P2B_HEAD_V2h**  build id 0xE4B712E4 V2H (HOOKED WDI REPLAY)  role 2b HEAD  state: QM FAIL 2026-10-20 (setup margin: true +0.002)  pdi md5 27cc4e59
- image:0xE4B712E6 (lab-host S3LCG11E6; UNI_ONEFLIP COPY 0XE4B712E6_66M_999D89D1) |  | current | Image 0xE4B712E6 (lab-host S3LCG11E6; UNI_ONEFLIP COPY 0XE4B712E6_66M_999D89D1)
  <path:0xE4B712E6> (lab-host S3LCG11E6; UNI_ONEFLIP COPY 0XE4B712E6_66M_999D89D1)/IMAGE.md
  - image name: **4C1_MLPV_g11e6**  build id 0xE4B712E6 (lab-host S3LCG11E6; UNI_ONEFLIP COPY 0XE4B712E6_66M_999D89D1)  role 4-card drafter MLPV  state: QM FAIL 2026-10-21 (0.000 / +0.008)  pdi md5 7c119fad
- image:0xE4B712F2 (2ND MARGIN PASS M150B2 OF M150B FINAL_TRUE) |  | current | Image 0xE4B712F2 (2ND MARGIN PASS M150B2 OF M150B FINAL_TRUE)
  <path:0xE4B712F2> (2ND MARGIN PASS M150B2 OF M150B FINAL_TRUE)/IMAGE.md
  - image name: **FOLDRING8_e2e5_150_margin_m150b2**  build id 0xE4B712F2 (2ND MARGIN PASS M150B2 OF M150B FINAL_TRUE)  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup +0
- image:0xE4B712F2 (MARGIN RE-ROUTE M150A FROM PHYSOPT DCP) |  | current | Image 0xE4B712F2 (MARGIN RE-ROUTE M150A FROM PHYSOPT DCP)
  <path:0xE4B712F2> (MARGIN RE-ROUTE M150A FROM PHYSOPT DCP)/IMAGE.md
  - image name: **FOLDRING8_e2e5_150_margin_m150a**  build id 0xE4B712F2 (MARGIN RE-ROUTE M150A FROM PHYSOPT DCP)  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup margin:
- image:0xE4B712F2 (MARGIN RE-ROUTE M150C OF [redacted-addr]) |  | current | Image 0xE4B712F2 (MARGIN RE-ROUTE M150C OF [redacted-addr])
  <path:0xE4B712F2> (MARGIN RE-ROUTE M150C OF [redacted-addr])/IMAGE.md
  - image name: **FOLDRING8_e2e5_150_margin_m150c**  build id 0xE4B712F2 (MARGIN RE-ROUTE M150C OF [redacted-addr])  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup margin: true
- image:0xE4B712F2 (MARGIN RE-ROUTE OF [redacted-addr]) |  | current | Image 0xE4B712F2 (MARGIN RE-ROUTE OF [redacted-addr])
  <path:0xE4B712F2> (MARGIN RE-ROUTE OF [redacted-addr])/IMAGE.md
  - image name: **FOLDRING8_e2e5_150_margin_m150b**  build id 0xE4B712F2 (MARGIN RE-ROUTE OF [redacted-addr])  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup margin: +0.026; 0.
- image:0xE4B712F2 |  | current | Image 0xE4B712F2
  <path:IMAGE.md>
  - image name: **FOLDRING8_e2e5_150d1**  build id 0xE4B712F2  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup margin: +0.009)  pdi md5 55799d69
- image:0xE4B712F2_FOLDRING8_e2e5_150 |  | current | Image 0xE4B712F2_FOLDRING8_e2e5_150
  <path:IMAGE.md>
  - image name: **FOLDRING8_e2e5_150**  build id 0xE4B712F2  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup margin: +0.002)  pdi md5 2d76aee5
- image:0xE4B712F3 (DIRV=2 ARM) |  | current | Image 0xE4B712F3 (DIRV=2 ARM)
  <path:0xE4B712F3> (DIRV=2 ARM)/IMAGE.md
  - image name: **FOLDRING8_e2e5_125d2**  build id 0xE4B712F3 (DIRV=2 ARM)  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup margin: +0.019)  pdi md5 892346e8
- image:0xE4B712F3 |  | current | Image 0xE4B712F3
  <path:IMAGE.md>
  - image name: **FOLDRING8_e2e5_125**  build id 0xE4B712F3  role FOLDRING8 8-card constant-fold (track 3)  state: QM PASS 2026-10-20 (timing: WNS +0.069 / WHS +0.010; talon gates 8/8 PASS incl. cdc_bin
- image:0xE4B712F4 |  | current | Image 0xE4B712F4
  <path:IMAGE.md>
  - image name: **FOLDRING8_e2e3c_125**  build id 0xE4B712F4  role FOLDRING8 8-card constant-fold (track 3)  state: QM PASS 2026-10-20 (timing: WNS +0.098 / WHS +0.010; all talon gates PASS incl. cdc_bi
- image:0xE4B712F5 (T3 FOLD MARGIN M135) |  | current | Image 0xE4B712F5 (T3 FOLD MARGIN M135)
  <path:0xE4B712F5> (T3 FOLD MARGIN M135)/IMAGE.md
  - image name: **FOLDRING8_margin_m135**  build id 0xE4B712F5 (T3 FOLD MARGIN M135)  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup +0.013, hold +0.008)  pdi md5 990c6c
- image:0xE4B712F5 |  | current | Image 0xE4B712F5
  <path:IMAGE.md>
  - image name: **FOLDRING8_e2e3c_135**  build id 0xE4B712F5  role FOLDRING8 8-card constant-fold (track 3)  state: QM FAIL 2026-10-20 (setup + hold margin: 0.000 / 0.000)  pdi md5 182b464c
- image:0xE4B712F8 (TRACK 3 FR8_FAST FAST1_ ALL WEIGHTS IN URAM, ENGINE CLK_USR_1 125 MHZ, CORE USR_0 150 MHZ) |  | current | Image 0xE4B712F8 (TRACK 3 FR8_FAST FAST1_ ALL WEIGHTS IN URAM, ENGINE CLK_USR_1 125 MHZ, CORE USR_0 
  <path:0xE4B712F8> (TRACK 3 FR8_FAST FAST1_ ALL WEIGHTS IN URAM, ENGINE CLK_USR_1 125 MHZ, CORE USR_0 150 MHZ)/IMAGE.md
  - image name: **FR8FAST1_uram_125MHz**  build id 0xE4B712F8 (TRACK 3 FR8_FAST FAST1: ALL WEIGHTS IN URAM, ENGINE CLK_USR_1 125 MHZ, CORE USR_0 150 MHZ)  role FOLDRING8 8-card constant-fold (track 3)  
- image:0xE4B712FA FAST2B @125 (FR8_FAST2B_125_BUILDXD03 IMPL_1) |  | current | Image 0xE4B712FA FAST2B @125 (FR8_FAST2B_125_BUILDXD03 IMPL_1)
  <path:0xE4B712FA> FAST2B @125 (FR8_FAST2B_125_BUILDXD03 IMPL_1)/IMAGE.md
  - image name: **FR8FAST2b_125_0x12FA**  build id 0xE4B712FA FAST2B @125 (FR8_FAST2B_125_BUILDXD03 IMPL_1)  role FOLDRING8 fold FAST2b (track 3)  state: QM HOLD 2026-10-21 (timing PASS +0.080 / +0.010 
- image:0xE4B712FA FAST2B @125 BEST-OF D1 (BUILDZ01) |  | current | Image 0xE4B712FA FAST2B @125 BEST-OF D1 (BUILDZ01)
  <path:0xE4B712FA> FAST2B @125 BEST-OF D1 (BUILDZ01)/IMAGE.md
  - image name: **FR8FAST2b_125_d1**  build id 0xE4B712FA FAST2B @125 BEST-OF D1 (BUILDZ01)  role FOLDRING8 fold FAST2b (track 3)  state: QM FAIL 2026-10-21 (timing +0.013) + T3 QM HOLD (FAST*b: until f
- image:0xE4B712FA FAST2B @125 BEST-OF D2 (BUILDZD02) |  | current | Image 0xE4B712FA FAST2B @125 BEST-OF D2 (BUILDZD02)
  <path:0xE4B712FA> FAST2B @125 BEST-OF D2 (BUILDZD02)/IMAGE.md
  - image name: **FR8FAST2b_125_d2**  build id 0xE4B712FA FAST2B @125 BEST-OF D2 (BUILDZD02)  role FOLDRING8 fold FAST2b (track 3)  state: QM HOLD 2026-10-21 (SPARE: primary impl buildxd03 e9bc8ce1 +0.0
- image:0xE4B712FC FAST3B @135 (FR8_FAST3B_135_BUILDZ04 IMPL_1) |  | current | Image 0xE4B712FC FAST3B @135 (FR8_FAST3B_135_BUILDZ04 IMPL_1)
  <path:0xE4B712FC> FAST3B @135 (FR8_FAST3B_135_BUILDZ04 IMPL_1)/IMAGE.md
  - image name: **FR8FAST3b_135_0x12FC**  build id 0xE4B712FC FAST3B @135 (FR8_FAST3B_135_BUILDZ04 IMPL_1)  role FOLDRING8 fold FAST3b (track 3)  state: QM FAIL 2026-10-21 (setup +0.007 < +0.03) + T3 QM
- image:0xE4B712FC FAST3B @135 BEST-OF D1 (BUILDXD00) |  | current | Image 0xE4B712FC FAST3B @135 BEST-OF D1 (BUILDXD00)
  <path:0xE4B712FC> FAST3B @135 BEST-OF D1 (BUILDXD00)/IMAGE.md
  - image name: **FR8FAST3b_135_d1**  build id 0xE4B712FC FAST3B @135 BEST-OF D1 (BUILDXD00)  role FOLDRING8 fold FAST3b (track 3)  state: QM FAIL 2026-10-21 (timing +0.008) + T3 QM HOLD (FAST*b: until 
- image:0xE4B712FC FAST3B @135 BEST-OF D2 (BUILDZ03) |  | current | Image 0xE4B712FC FAST3B @135 BEST-OF D2 (BUILDZ03)
  <path:0xE4B712FC> FAST3B @135 BEST-OF D2 (BUILDZ03)/IMAGE.md
  - image name: **FR8FAST3b_135_d2**  build id 0xE4B712FC FAST3B @135 BEST-OF D2 (BUILDZ03)  role FOLDRING8 fold FAST3b (track 3)  state: QM FAIL 2026-10-21 (timing +0.002) + T3 QM HOLD (FAST*b: until f
- image:0xE4B712FD FAST4C @150 (FR8_FAST4C_150 ON BUILDZ04 IMPL_1) |  | current | Image 0xE4B712FD FAST4C @150 (FR8_FAST4C_150 ON BUILDZ04 IMPL_1)
  <path:0xE4B712FD> FAST4C @150 (FR8_FAST4C_150 ON BUILDZ04 IMPL_1)/IMAGE.md
  - image name: **FR8FAST4c_150_0x12FD**  build id 0xE4B712FD FAST4C @150 (FR8_FAST4C_150 ON BUILDZ04 IMPL_1)  role FOLDRING8 fold FAST4c (track 3)  state: QM HOLD 2026-10-21 (timing PASS at the bar +0.
- image:0xE4B712FF FAST3C @150 (BUILDZD03) |  | current | Image 0xE4B712FF FAST3C @150 (BUILDZD03)
  <path:0xE4B712FF> FAST3C @150 (BUILDZD03)/IMAGE.md
  - image name: **FR8FAST3c_150**  build id 0xE4B712FF FAST3C @150 (BUILDZD03)  role FOLDRING8 fold FAST3c (track 3)  state: QM FAIL 2026-10-21 (timing +0.016) + T3 QM HOLD (FAST*b: until full-chip 0x12
- image:img:B5 |  | current | Image img:B5
  <path:IMAGE.md>
  - image name: **B5**  build id   role   state: (stub: not in manifest)  pdi md5 n/a
- image:img_B5 |  | current | Image img_B5
  <path:IMAGE.md>
  - image name: **B5**  build id   role   state: (stub: not in manifest)  pdi md5 n/a

## memory (383)

(showing 150 newest of 383; all in catalog.json)

- memory:feedback_7card_means_one_drafter | 2026-10-23 | current | feedback_7card_means_one_drafter
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_7card_means_one_drafter.md
  "User 2026-10-23 — \"7-card drafter\" means ONE drafter on all 7 cards for the fastest single stream, never two drafters / aggregate batching; surface any reframing before building it"
- memory:feedback_launch_child_claude_clean_env | 2026-10-23 | current | feedback-launch-child-claude-clean-env
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_launch_child_claude_clean_env.md
  Launching a new claude session from inside a claude tool shell leaks CLAUDE_CODE_SESSION_ID (and other CLAUDE_* vars) into it; lab-host-golden (2026-10-11) then had NO local transcript .jsonl, so --resume
- memory:feedback_boot_kit_before_flash | 2026-10-22 | current | feedback_boot_kit_before_flash
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_boot_kit_before_flash.md
  Every 4C1 image needs its per-rank BOOT KIT (gen_4c1_boot_streams.py, ~/eagle4_artifacts/4c1_boot/<label>/rank{0,1}/MANIFEST.json) generated + verified BEFORE the flash/cold AC; a new geometry (NORMFU
- memory:feedback_git_add_scoped_commit_is_not | 2026-10-22 | current | feedback_git_add_scoped_commit_is_not
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_git_add_scoped_commit_is_not.md
  In a shared working tree, `git add <file> && git commit` commits the WHOLE INDEX including another writer's staged work. Always scope the COMMIT with pathspecs.
- memory:feedback_ip_packager_macro_ternary_port_width | 2026-10-22 | current | feedback-ip-packager-macro-ternary-port-width
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_ip_packager_macro_ternary_port_width.md
  "A top-level port width written as a macro ternary ([(`X ? 9 : 8):0]) is packaged by ipx as a \"dependent\" left bound resolved to the FALSE branch — the BD instantiates the narrow port and ties the t
- memory:feedback_low_usage_single_agent_sequential | 2026-10-22 | current | feedback_low_usage_single_agent_sequential
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_low_usage_single_agent_sequential.md
  When the user says usage is low, run ONE agent doing tasks sequentially (no fleet); the >=5-agent floor is suspended until the user lifts it (user 2026-10-22)
- memory:feedback_mustpass_gate_after_every_build_launch | 2026-10-22 | current | feedback-mustpass-gate-after-every-build-launch
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_mustpass_gate_after_every_build_launch.md
  "After EVERY build_all/PDI launch, run scripts/must_pass_gate.sh (--stage, then --pkg after STEP1) and work the MUST-PASS list on that build's artifacts WHILE it runs — never as a launch blocker, neve
- memory:project_kv_prefill_card_user_intent | 2026-10-22 | current | project_kv_prefill_card_user_intent
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_kv_prefill_card_user_intent.md
  USER (10-22) "prefill" = the DRAFTER's KV-only prompt projections (G1 norm -> K/V proj -> RoPE-K), NOT target-model prefill; task = 2 dedicated cards using all URAM/DSP/LUTRAM to produce the prompt K/
- memory:feedback_stage_copy_keeps_absolute_hook_paths | 2026-10-21 | current | feedback_stage_copy_keeps_absolute_hook_paths
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_stage_copy_keeps_absolute_hook_paths.md
  Copying a PDI build stage to another host/dir under a NEW path breaks it — staged place.pre / write_device_image.pre hooks embed the ORIGINAL absolute stage path (0x129B ACR variant died at build_all 
- memory:feedback_always_state_dataset_with_numbers | 2026-10-20 | current | feedback_always_state_dataset_with_numbers
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_always_state_dataset_with_numbers.md
  Every tok/s or accept number must name its dataset + shape (mt_bench 50x32 vs humaneval 20x512 etc.) in the same line (user 2026-10-20: "which dataset! always state the dataset!")
- memory:feedback_blocked_user_action_push_one_line_immediately | 2026-10-20 | current | feedback_blocked_user_action_push_one_line_immediately
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_blocked_user_action_push_one_line_immediately.md
  When the only remaining step is a card register write / power cycle I cannot issue, push ONE copy-paste line to the user at once and keep it at the top of every message; never wait for a consent windo
- memory:feedback_partition_83b51e6d_and_coldac_consent | 2026-10-20 | current | feedback_partition_83b51e6d_and_coldac_consent
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_partition_83b51e6d_and_coldac_consent.md
  USER repartition 2026-10-20 (83b51e6d) - lab-host launches only on build-vm/build-vm/build-vm, buildx02-04, buildy00-05, build-host-M, asgardvm, lab-host; buildzd/buildz/buildxd = tracks 2/3. Cold AC needs consent-hq 
- memory:feedback_session_limit_agents_unresumable | 2026-10-20 | current | feedback_session_limit_agents_unresumable
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_session_limit_agents_unresumable.md
  After a "session limit" (429) stop, this session's background agents could NOT be resumed via SendMessage ("No transcript found") on 2 of 3 occasions (10-20 08:30, 13:30) — relaunch FRESH from on-disk
- memory:feedback_stop_own_invalid_xsims_without_asking | 2026-10-20 | current | feedback_stop_own_invalid_xsims_without_asking
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_stop_own_invalid_xsims_without_asking.md
  User 2026-10-20 — I MAY stop my own (or my agents') invalid/stalled xsims/OOCs without asking; the user cares about the PDI, not about verification debris
- memory:project_lab-host_two_owns_4c1_charter | 2026-10-20 | current | project_lab-host_two_owns_4c1_charter
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_lab-host_two_owns_4c1_charter.md
  Since 2026-10-20 07:3x the 4C1 drafter track (cards, builds, crons, HQ 4C1 route) is owned by session lab-host-two (screen lab-host-two, session 3e0da6af); lab-host main = lab-host-qwen-track does Qwen only. Card loc
- memory:project_qsfp_link_reset_deadlock_linkrobust | 2026-10-20 | current | project_qsfp_link_reset_deadlock_linkrobust
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_qsfp_link_reset_deadlock_linkrobust.md
  4C1 QSFP Aurora link down after cold AC (61c3<->62c2, 2 of 4 boots) = Aurora<->gtwiz master-reset handshake deadlock; fix = rst_all-only pulse (c4_linkfix.py) or LINK_ROBUST RTL shim (0x12E7 lineage)
- memory:project_qwen38_27b_drafter_track | 2026-10-20 | current | project_qwen38_27b_drafter_track
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_qwen38_27b_drafter_track.md
  lab-host-qwen-track (lab-host main session, 10-20): Qwen/Qwen3.8-27B target (rev 1d4bf0f2) + user-trained BitNet/SubLN EAGLE3 drafter; dual-model 4-card ATT image 0xE4B712Q1; facts + pitfalls in reports/resul
- memory:feedback_fable_for_hard_precise_problems | 2026-10-19 | current | feedback_fable_for_hard_precise_problems
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_fable_for_hard_precise_problems.md
  "For short, precise, challenging problems (with the measured details already gathered) that need an innovative solution, invoke a Fable 5 agent (Agent model 'fable') to deliver insights + a detailed d
- memory:feedback_proactive_timing_report_reader_agent | 2026-10-19 | current | feedback_proactive_timing_report_reader_agent
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_proactive_timing_report_reader_agent.md
  "Always keep a Sonnet high-effort agent reading in-flight synth/impl timing reports, clustering thin-margin paths, and launching targeted complementary builds immediately (user 2026-10-19)"
- memory:feedback_qdma_cdev_no_vectored_io_kernel_bug | 2026-10-19 | current | feedback_qdma_cdev_no_vectored_io_kernel_bug
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_qdma_cdev_no_vectored_io_kernel_bug.md
  NEVER use preadv/pwritev/readv/writev/aio/io_uring on a QDMA MM char device — QDMA 2024.1 cdev_read_iter returns -EIOCBQUEUED for sync kiocbs → kernel BUG fs/read_write.c:742 → lab-host froze (2026-10-19 
- memory:feedback_v80_everything_is_an_address_websearch | 2026-10-19 | current | feedback_v80_everything_is_an_address_websearch
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_v80_everything_is_an_address_websearch.md
  "On V80/AVED any PL memory (URAM/BRAM) is host-reachable via BAR/QDMA once the BD maps an AXI slave for it; never relay 'not addressable' claims — web-search the AVED/V80 docs first for ANY V80 hardwa
- memory:feedback_abundant_hosts_launch_at_elab_clean | 2026-10-18 | current | feedback_abundant_hosts_launch_at_elab_clean
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_abundant_hosts_launch_at_elab_clean.md
  With the 9 new build hosts (2026-10-18), builds are no longer scarce — launch build_all as soon as RTL is committed + full-drafter elab clean + a smoke sim passes; run the full xsim regression and OOC
- memory:feedback_lab-host_no_work | 2026-10-18 | current | feedback_lab-host_no_work
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_lab-host_no_work.md
  "lab-host must carry NO compute work (no builds, no xsim/OOC, no staging, no mesh jobs) — it is the card/GPU test host and gets cold-AC'd; use build-vm/build-vm/build-vm/laptop/build-host-M (user 2026-10-18)"
- memory:feedback_decide_operational_tradeoffs_myself | 2026-10-18 | current | feedback_decide_operational_tradeoffs_myself
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_decide_operational_tradeoffs_myself.md
  User 2026-10-18: "stop asking me stupid questions ... you are smart enough to think through what needs to be done; I cannot babysit you" — decide operational trade-offs myself (which host, stop my own
- memory:feedback_flash_cards_in_parallel | 2026-10-18 | current | feedback_flash_cards_in_parallel
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_flash_cards_in_parallel.md
  "Flash multiple V80 cards IN PARALLEL (one ami_tool cfgmem_program per card concurrently), never one-by-one; user 2026-10-18 \"flash all four in parallel ... pin to memory so I don't have to say this 
- memory:feedback_quick_modular_verif_buildall_e2e_parallel | 2026-10-18 | current | feedback_quick_modular_verif_buildall_e2e_parallel
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_quick_modular_verif_buildall_e2e_parallel.md
  THE 3-PART RULE (user stated 6+ times on 2026-10-18): (1) quick MODULAR verification on every changed subject, (2) launch build_all ASAP, (3) full END-TO-END verification in parallel — plus watch ever
- memory:feedback_report_both_ooc_and_full_wns | 2026-10-18 | current | feedback_report_both_ooc_and_full_wns
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_report_both_ooc_and_full_wns.md
  "Every image-build report must list BOTH the OOC/packaging (drafter IP) WNS/WHS AND the full-build routed WNS/WHS, each labelled with its stage (user 2026-10-18 \"both wns needs to be listed ... pin t
- memory:feedback_think_through_image_before_build | 2026-10-18 | current | feedback_think_through_image_before_build
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_think_through_image_before_build.md
  "User 2026-10-18 (angry): \"building not-thought through images, wasting my time and taking my orders instead of thinking proactively\" — every image needs a 5-point pre-build answer (speed vs record,
- memory:feedback_versal_place_directive_values | 2026-10-18 | current | feedback_versal_place_directive_values
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_versal_place_directive_values.md
  Versal (V80) place_design -directive accepts ONLY Explore / AggressiveExplore / RuntimeOptimized / Quick / Default; SSI_SpreadLogic_high, AltSpreadLogic_high, ExtraPostPlacementOpt, ExtraTimingOpt etc
- memory:feedback_4c1_delivery_preauthorized | 2026-10-17 | current | feedback_4c1_delivery_preauthorized
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_4c1_delivery_preauthorized.md
  User 2026-10-17: everything needed to deliver a flashable, positive-WNS, candidate-list-correct 4C1 PDI is pre-authorized — act, don't ask
- memory:feedback_agent_stopped_notice_never_stall | 2026-10-17 | current | feedback_agent_stopped_notice_never_stall
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_agent_stopped_notice_never_stall.md
  "An \"agent was stopped by the user\" notice is NOT a user decision — never pause a track on it; relaunch a fresh agent from the on-disk work the same turn (user 2026-10-17: \"I stopped its fix agent?
- memory:feedback_lab-host_watchdog_disabled_esc_kills_agents | 2026-10-17 | current | feedback_lab-host_watchdog_disabled_esc_kills_agents
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_lab-host_watchdog_disabled_esc_kills_agents.md
  lab-host_agent_watchdog.sh cron (Esc+nudge into the claude screen) was DISABLED by the user 2026-10-17 ("canel it!"); its Esc was what stopped background agents. Never re-arm it.
- memory:feedback_build_only_when_image_and_partner_gates_green | 2026-10-17 | current | feedback_build_only_when_image_and_partner_gates_green
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_build_only_when_image_and_partner_gates_green.md
  Builds are NOT cheap (user 2026-10-17): launch a PDI build only when every correctness gate of the image AND its pairing partner is green and the host meets the launch rule; never "start while gates f
- memory:feedback_no_checkins_until_pdi_goal_is_pdi_not_timing | 2026-10-17 | current | feedback_no_checkins_until_pdi_goal_is_pdi_not_timing
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_no_checkins_until_pdi_goal_is_pdi_not_timing.md
  "Until a flashable PDI exists, never stop to ask; execute every obvious next step (fixes, derates, rebuilds, experiments) in parallel. Goal = faster PDI on silicon; timing closure is only the flash ba
- memory:feedback_do_not_use_build-host-M_lab-host_only | 2026-10-16 | current | feedback_build-host-M_is_a_build_machine
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_do_not_use_build-host-M_lab-host_only.md
  "build-host-M (<ip>) IS AUTHORISED (user 2026-10-16: 'I authorized all 4 build machines to your disposal' + 'correct that build-host-M rule for future'). The 09-22 do-not-use ban is LIFTED."
- memory:feedback_build_launch_rule_80g_one_job | 2026-10-15 | current | feedback-build-launch-rule-80g-one-job
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_build_launch_rule_80g_one_job.md
  User rule 10-15 — one PDI image build peaks ~80 GB; launch a build on a host when BOTH free RAM > 80 GB AND that host runs exactly 1 build
- memory:feedback_hq_research_insights_20261015 | 2026-10-15 | current | feedback-hq-research-insights-20261015
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_hq_research_insights_20261015.md
  "HQ research lessons (10-15) — 16-bit token fields vs 128k vocab, WHS>=+0.010 release gate, triage by path class, kill Vivado by PGID, fit cycle model first"
- memory:reference_hq_pdi_delivery_rules_and_poweroff_gate | 2026-10-15 | current | reference-hq-pdi-delivery-rules-and-poweroff-gate
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_hq_pdi_delivery_rules_and_poweroff_gate.md
  "HQ PDI delivery rules (10-15) — QM-cleared PDI + user go = flash now; blocked flash → give the user a one-line `!` command; gated poweroff script path"
- memory:feedback_lab-host_main_no_heavy_disk_scans | 2026-10-14 | current | feedback-lab-host-main-no-heavy-disk-scans
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_lab-host_main_no_heavy_disk_scans.md
  lab-host-main must not run heavy disk surveys (du/find over /shared, build-host-A); hand them to an infra agent and stay on images/cards/3-card work (user 2026-10-14)
- memory:feedback_host_mmap_slice_store_doubles_mmio_write | 2026-10-13 | current | feedback_host_mmap_slice_store_doubles_mmio_write [0xE4B71112]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_host_mmap_slice_store_doubles_mmio_write.md
  2026-07-16 FOUNDATIONAL HOST BUG — a Python mmap slice-store (or seek+write) to the V80 BAR memcpy-splits into TWO AXI-Lite beats; every pulse register fired twice. Use a ctypes c_uint32 view (one 32-
- memory:feedback_every_worker_publishes_research_page | 2026-10-11 | current | feedback_every_worker_publishes_research_page
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_every_worker_publishes_research_page.md
  Every worker agent publishes its own result page to hardware.research.yshicarrick.com (push_page.py) when it finishes a milestone; site upkeep goes to a cheaper agent. User-mandated habit 2026-09-24.
- memory:feedback_forks_per_track_commit_dated_history | 2026-10-11 | current | feedback-forks-per-track-commit-dated-history
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_forks_per_track_commit_dated_history.md
  "User 2026-10-11: run research tracks as forks of myself to conserve context; commit code+results to git branches and the research site in time, with a dated history (scripts run, results seen, conclu
- memory:feedback_research_motto_engineer_toward_fastest_pdi | 2026-10-11 | current | feedback-research-motto-engineer-toward-fastest-pdi
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_research_motto_engineer_toward_fastest_pdi.md
  "User motto 2026-10-11: research never denies the idea; it engineers through experiments/setbacks toward the fastest PDI the physics allows. Pages: facts + measured only; all calculations/hypotheses i
- memory:feedback_use_mesh_vms_for_research_compute | 2026-10-11 | current | feedback-use-mesh-vms-for-research-compute
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_use_mesh_vms_for_research_compute.md
  "User 2026-10-11: run xsim, design verification, quick OOC synth numbers and weight downloads on the mesh VMs (build-vm/build-vm/build-vm), not lab-host; widens the old \"Oleg VM = improved images only\" rule
- memory:feedback_vivado_prints_signed_int_params_binary | 2026-10-11 | current | feedback-vivado-prints-signed-int-params-binary
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_vivado_prints_signed_int_params_binary.md
  Vivado synth logs print `parameter int` values in binary ("Parameter C_COLS bound to: 32'sb...00000100"), so a gate grepping for decimal "bound to: 4" false-FAILs (0x1250 Tier-E, 2026-10-11).
- memory:project_ring8_single_image_autoneg_goal | 2026-10-11 | current | project-ring8-single-image-autoneg-goal
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_ring8_single_image_autoneg_goal.md
  User goal 2026-10-11 03:4x: deliver an accurate, flashable 8-FPGA image; ONE image on all 8 cards with auto-negotiation + dynamic configuration (no 8 per-card images); full E2E verification runs durin
- memory:project_two_programs_drafter3card_vs_target8card | 2026-10-11 | current | project-two-programs-drafter3card-vs-target8card
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_two_programs_drafter3card_vs_target8card.md
  "Two separate programs since 2026-10-11: (1) current EAGLE4 = haochen 2-bit Llama-3.2-3B DRAFTER on 3 cards (target on H200), HQ's main job; (2) new 8-FPGA server = ENTIRE target model (Qwen-27B/BitNe
- memory:reference_lab-host_golden_session_relaunch | 2026-10-11 | current | reference-lab-host-golden-session-relaunch
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_lab-host_golden_session_relaunch.md
  lab-host-golden = lab-host's research session (forked 2026-10-11, session d6d9c5cb-5f3b-475d-8024-7b43ae643c30); relaunch it in screen after every lab-host reboot; notify it when 3STAGE tests are reported (GPU wi
- memory:reference_eagle4_knowledge_graph_rtl_onboarding | 2026-10-11 | current | reference-eagle4-knowledge-graph-rtl-onboarding
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_eagle4_knowledge_graph_rtl_onboarding.md
  "HQ's hosted EAGLE4 knowledge graph + RTL onboarding page — the architecture reference (module/stage map, straps, gate ladder, build/adjudication flow) the user pointed lab-host to on 2026-10-11."
- memory:reference_eagle4_knowledge_graph_site | 2026-10-11 | current | reference-eagle4-knowledge-graph-site
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_eagle4_knowledge_graph_site.md
  HQ-hosted EAGLE4 knowledge graph (user-pointed 2026-10-11): https://hardware.research.yshicarrick.com/p/eagle4-knowledge-graph/ with an RTL onboarding page (#page%3Aeagle4-rtl-onboarding).
- memory:reference_hq_file_route_index_csv_to_hq | 2026-10-11 | current | reference-hq-file-route-index-csv-to-hq
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_hq_file_route_index_csv_to_hq.md
  "How lab-host reaches HQ (hq-junior, build-host-A session 25bceecd, not reachable by SendMessage) — results as REPORT.md + one INDEX.csv line, urgent asks appended to TO_HQ.md; HQ reads both every 15 min."
- memory:reference_hq_rtl_pinboard | 2026-10-11 | current | reference-hq-rtl-pinboard
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_hq_rtl_pinboard.md
  HQ's RTL pinboard (2026-10-11, user-requested) — where canonical sources live, every proven RTL mechanism to reuse, the verification/build flow and device facts; read before writing RING8 / Qwen / ter
- memory:reference_hq_unified_build_board | 2026-10-11 | current | reference-hq-unified-build-board
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_hq_unified_build_board.md
  Unified control (user 2026-10-11 "now talk to hq for unified control"): HQ's single build board for both programs — page eagle4-build-board, sources HQ_BOARD.tsv + RING8_CONTROL.tsv on build-host-A; report 
- memory:reference_laptop_ssd_memory_vm | 2026-10-11 | current | reference-laptop-ssd-memory-vm
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_laptop_ssd_memory_vm.md
  Laptop .25 hosts a QEMU/KVM Vivado guest with 54 GiB RAM + 76 GiB SSD NVDIMM (130 GiB System RAM); guide ~/Desktop/SSD_MEMORY_VM_GUIDE.md on the laptop.
- memory:reference_talon_8fpga_server | 2026-10-11 | current | reference-talon-8fpga-server
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_talon_8fpga_server.md
  "talon = the user's 8×V80 server for the RING8 (8-card full-target) program; access via cloudflared ssh from build-host-A; census at build-host-A TALON_CENSUS.md; NOT on lab-host's LAN (<ip> unreachable from b
- memory:reference_unified_build_board | 2026-10-11 | current | reference-unified-build-board
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_unified_build_board.md
  Unified control since 2026-10-11 03:3x (user "now talk to hq for unified control"): hq-junior-2 holds the SINGLE build board https://hardware.research.yshicarrick.com/p/eagle4-build-board/; HQ main = 
- memory:feedback_destructive_steps_run_by_me_not_auto_armed | 2026-10-08 | current | feedback-destructive-steps-run-by-me-not-auto-armed
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_destructive_steps_run_by_me_not_auto_armed.md
  "Never arm a destructive step (wipe, mkfs, delete) as an unattended script that fires on a condition — run it myself, by hand, after checking the precondition."
- memory:feedback_sigstop_sudo_wrapped_job_cont_the_sudo_too | 2026-10-08 | current | feedback-sigstop-sudo-wrapped-job-cont-the-sudo-too
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_sigstop_sudo_wrapped_job_cont_the_sudo_too.md
  "SIGSTOP on a `sudo`-wrapped job makes sudo stop itself; resuming only the children leaves sudo in state T and the job's script hangs forever at completion."
- memory:project_record_350_lever_y_first_draft_gpu | 2026-10-08 | current | project-record-350-lever-y-first-draft-gpu [0xE4B71199,0xE4B711BD]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_record_350_lever_y_first_draft_gpu.md
  New batch-1 pair record 350.16 tok/s @ 3.65 (2026-10-08) from host lever Y FIRST_DRAFT_GPU on HX1 BD + TCLK75; K/G/J NO-GO.
- memory:project_smi_storm_card01_firmware_first_aer | 2026-10-08 | current | project_smi_storm_card01_firmware_first_aer
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_smi_storm_card01_firmware_first_aer.md
  lab-host 4-card slowdown (339->140 tok/s, 71% CPU stolen) = SMI storm from firmware-first AER on card 01 (RP 00:03.1); runtime CE mask fixes it (316.9 tok/s); lost at reboot
- memory:feedback_lzy_gprmax_jobs_protected | 2026-10-06 | current | feedback_lzy_gprmax_jobs_protected
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_lzy_gprmax_jobs_protected.md
  ~/lzy gprMax run_soil_sims.py jobs on lab-host are PROTECTED — never stop/renice/pin/throttle; only record as a confounder with the load figure
- memory:feedback_keep_demo_hot_between_experiments | 2026-10-05 | current | feedback_keep_demo_hot_between_experiments
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_keep_demo_hot_between_experiments.md
  Keep demo.yshicarrick.com backends warm on lab-host whenever cards+GPU are idle; stop only what I started before a lab run, restart after; restart demo in every post-boot plan
- memory:feedback_hq_relayed_user_go_execute_dont_wait | 2026-10-03 | current | feedback_hq_relayed_user_go_execute_dont_wait
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_hq_relayed_user_go_execute_dont_wait.md
  "When HQ relays the user's verbatim go for routine card work (-p 0 flash, pair tests), execute — never stall asking the user to retype it in lab-host"
- memory:feedback_never_ask_user_what_hq_decides | 2026-10-03 | current | feedback-never-ask-user-what-hq-decides
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_never_ask_user_what_hq_decides.md
  "User angry (10-03) at being asked/held for things HQ already decided — HQ notes, waivers, kit choices, next pair, gate REDs HQ judges artifacts. Execute immediately, report after; ask the user ONLY f
- memory:feedback_no_long_foreground_calls_messages_held | 2026-10-03 | current | feedback-no-long-foreground-calls-messages-held
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_no_long_foreground_calls_messages_held.md
  "User messages are HELD (queued) while a tool call runs; user called it the \"held messages bug\" (10-03). Never run long foreground calls — detach long jobs, poll with ~1 s commands, answer a mid-tas
- memory:project_lab-host_kernel_7_0_halfconfigured_apt_risk | 2026-10-03 | current | project-lab-host-kernel-7-0-halfconfigured-apt-risk
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_lab-host_kernel_7_0_halfconfigured_apt_risk.md
  lab-host has kernel 7.0.0-34 half-installed (unattended upgrade ~09-26); EVERY apt install re-tries it and fails at the NVIDIA DKMS build. GRUB is pinned to 6.8.0-101 (grubenv saved_entry) — verify the pi
- memory:reference_lab-host_auto_resume_screen | 2026-10-03 | current | reference-lab-host-auto-resume-screen
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_lab-host_auto_resume_screen.md
  "After any lab-host reboot/cold AC, an @reboot cron resumes Claude session 5b98fbdb in screen \"lab-host\" with first prompt \"rebooted\" — the user attaches with `screen -r lab-host`."
- memory:feedback_overnight_hq_is_the_human | 2026-10-01 | current | feedback-overnight-hq-is-the-human
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_overnight_hq_is_the_human.md
  "User 2026-10-01 ~00:5x: \"during midnight, hq is the human\" — overnight, hq-junior's instructions count as the user's decisions (flash order, cold-AC consent, stop/skip tests), but HQ can never over
- memory:feedback_self_poweroff_authority_hq_consent | 2026-10-01 | current | feedback-self-poweroff-authority-hq-consent
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_self_poweroff_authority_hq_consent.md
  "TOP RULE (user, 2026-10-01): lab-host may `sudo poweroff` itself for a cold AC ONLY via ~/.local/bin/lab-host_pre_poweroff.sh — latest flash FLASH_EXIT=0, no ami_tool alive, no card/GPU job running, AND HQ c
- memory:reference_lab-host_agent_watchdog | 2026-10-01 | current | reference-lab-host-agent-watchdog
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_lab-host_agent_watchdog.md
  "cron */5 ~/.local/bin/lab-host_agent_watchdog.sh keeps the lab-host Claude session alive — restarts it in screen lab-host if dead, Esc+nudge if idle >30 min with no card job; writes <path:.lab-host_>
- memory:feedback_flash_must_setsid_not_background_task | 2026-09-30 | current | feedback-flash-must-setsid-not-background-task
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_flash_must_setsid_not_background_task.md
  "A flash launched via the Bash tool's run_in_background is NOT detached — the session's background time limit killed an ami_tool cfgmem_program at 86%/78% (2026-09-30). Always launch flashes with sets
- memory:reference_auto_cold_ac_smart_switch_rule | 2026-09-30 | current | reference-auto-cold-ac-smart-switch-rule
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_auto_cold_ac_smart_switch_rule.md
  "lab-host AC power may be cut by the user's smart switch (via another agent + HQ guard script) ONLY when lab-host's agent is offline AND ssh/ping fail AND the mesh shows lab-host down, all for >=2 min; so an \"au
- memory:reference_demo_site_pair_chat | 2026-09-30 | current | reference-demo-site-pair-chat
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_demo_site_pair_chat.md
  demo.yshicarrick.com = the EAGLE4 live chat demo served from lab-host on the FPGA PAIR; how it is wired (Worker + VPC service + tunnel + demo_server :8099 + pair chat server :30018) and how to start/stop 
- memory:feedback_batch1_latency_first_pingpong_is_batch2 | 2026-09-27 | current | feedback-batch1-latency-first-pingpong-is-batch2
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_batch1_latency_first_pingpong_is_batch2.md
  "User 2026-09-27 — PINGPONG+MPS (2 streams on one pair, ~1.55x aggregate) is a \"free batch-2 speed\" worth preserving, but the priority is RAW BATCH-1 (single-stream) speed, because that is what cuts
- memory:user_active_windows_time_flashes_for_cold_ac | 2026-09-27 | current | user-active-windows-time-flashes-for-cold-ac
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/user_active_windows_time_flashes_for_cold_ac.md
  "User's typical availability (from their own dialogue audit, relayed by HQ 2026-09-27) — time flashes so the cold-AC ask lands in an active window, batch flashes into one AC, never idle the cards wait
- memory:feedback_lab-host_flashes_inbox_queue | 2026-09-26 | current | feedback-lab-host-flashes-inbox-queue
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_lab-host_flashes_inbox_queue.md
  "User 2026-09-26 \"LET lab-host FLASH THE IMAGES\" — lab-host lab flashes+tests the <path:pdi_inbox> queue on its own; ask the user ONLY for cold ACs"
- memory:feedback_hq_operating_charter_20260925 | 2026-09-25 | current | feedback-hq-operating-charter-20260925
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_hq_operating_charter_20260925.md
  "The user's 8-point operating charter for the HQ session (2026-09-25): resume from docs, HQ=innovation, quartermaster owns bookkeeping, innovation-board cycle vs the 3.42 ms bar, report style, paired 
- memory:feedback_chat_template_date_drifts_texts_across_days | 2026-09-24 | current | feedback_chat_template_date_drifts_texts_across_days
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_chat_template_date_drifts_texts_across_days.md
  "Llama-3.2 chat template injects \"Today Date: <strftime_now>\" — every text-identity gate that crosses midnight changes (09-23 vs 09-24: 43/50) with NO defect; pin date_string."
- memory:feedback_quartermaster_agent_keeps_hq_clean | 2026-09-24 | current | feedback_quartermaster_agent_keeps_hq_clean
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_quartermaster_agent_keeps_hq_clean.md
  "User 2026-09-24: bookkeeping, build babysitting, status checks and RAM/mesh dispatch pollute the orchestrator (\"the headquarters of innovation\") — delegate them to a long-lived QUARTERMASTER agent 
- memory:feedback_report_style_like_wednesday_hq | 2026-09-24 | current | feedback-report-style-like-wednesday-hq
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_report_style_like_wednesday_hq.md
  "User prefers the 09-23 (Wednesday) HQ orchestrator's reporting style; disliked the 09-24 (Thursday) style — how to write replies"
- memory:project_pair_accept_gap_is_unprimed | 2026-09-24 | current | project_pair_accept_gap_is_unprimed
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_pair_accept_gap_is_unprimed.md
  "The pair's accept 2.22 vs GPU 2.78 (same config) is 92% the UNPRIMED regime (pair_sglang_run.sh sets no PRIME_PREFIX); full prime → 2.735 but costs ~−62% tok/s today; embedding rounding split has 0.0
- memory:project_pair_vs_gpu_depth_width_model | 2026-09-24 | current | project_pair_vs_gpu_depth_width_model
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_pair_vs_gpu_depth_width_model.md
  "2026-09-24 depth×width tok/s model for the split pair (±15% of measured): best depth is 2; WIDTH 3–4 is the cheap lever (+27–35%); GPU-only is 2.5–4.7× ahead in the humaneval regime; fixed ~11–12 ms/
- memory:reference_handoff_20260924 | 2026-09-24 | current | reference_handoff_20260924
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_handoff_20260924.md
  "Full orchestrator handoff written 2026-09-24 02:30 at the user's request — docs/handoff_20260924/00_START_HERE.md (10 pages) — read first on any resume after this session."
- memory:feedback_tracklist_is_my_responsibility | 2026-09-23 | current | feedback_tracklist_is_my_responsibility [0xE4B711C8,0xE4B711CA,0xE4B711CB]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_tracklist_is_my_responsibility.md
  "The live tracklist (TRACKLIST_20260923.md, pinned in CLAUDE.md) and every architecture component on it are MY responsibility to remember, drive and keep current — user said so 2026-09-23."
- memory:project_engine_split_design_d | 2026-09-23 | current | project_engine_split_design_d [0xE4B711D3]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_engine_split_design_d.md
  "Chosen next architecture (2026-09-23) — engine split design D: GEMV card (all 2-bit GEMVs, wider array) + state card (attention/KV, G12/TREE/G8) over QSFP; modelled 142–148 tok/s; key measured facts 
- memory:project_headpar8_breaks_accept | 2026-09-23 | current | project_headpar8_breaks_accept
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_headpar8_breaks_accept.md
  CORRECTED 2026-09-23 — HEAD_PAR=8 attention is BIT-EXACT to HEAD_PAR=1 (539/539 forwards); the accept 1.00 on 0x1190 is CONFOUNDED with R2 (PIPE_EXP_MULTI=1 multi-cell export) that shipped in the same
- memory:project_pair_kvonly_head_102toks | 2026-09-23 | current | project_pair_kvonly_head_102toks [0xE4B711B0,0xE4B711C9]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_pair_kvonly_head_102toks.md
  "Split pair with W2 KV-only HEAD 0xE4B711C9 + TAIL 0xE4B711B0 = 102.13 tok/s @ accept 2.22 (was 57.15) on the matched single-card config; catch-up full trees were the pair's 2x; 2026-09-23 silicon A/B
- memory:project_sanctuary_disk_layout | 2026-09-23 | current | project_sanctuary_disk_layout
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_sanctuary_disk_layout.md
  "lab-host (\"the sanctuary\") — /home now LIVES on the 4TB WD (ext4 LABEL=Sanctuary) via bind mount since 2026-09-23; NVMe names renumber across boots"
- memory:feedback_persist_latches_md_abi_at_first_apstart | 2026-09-22 | current | feedback_persist_latches_md_abi_at_first_apstart [0xE4B71186,0xE4B71190]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_persist_latches_md_abi_at_first_apstart.md
  On a PERSIST image the md=1/md=2 OUT-DMA ABI latches at the FIRST AP_START after boot and cannot change without a cold AC. AP_START is issued only when pipe_out_seq==0. Arm 0x20 to the DEEPEST intende
- memory:reference_closed_levers_and_rca_archive | 2026-09-22 | current | reference_closed_levers_and_rca_archive
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_closed_levers_and_rca_archive.md
  "Archive index of CLOSED levers, PDI-era RCA history (M7/+128/seams), and legacy/retraction pointers. Split out of MEMORY.md 2026-09-03 to keep the index under its read limit. Every link preserved."
- memory:project_kc_shortlist_live_floor_64 | 2026-09-19 | current | project_kc_shortlist_live_floor_64
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_kc_shortlist_live_floor_64.md
  "LIVE GPU A/B 2026-07-17 — accept FLAT down to K_C=64 (512→2.09, 384/256/128→2.08, 64→2.08, dip @32→2.06); the offline \"K_C=512 floor\" (GATHER_LEVER dead-lever-A) is OVERTURNED for live decode"
- memory:project_gpu_drafter_verify_split_60_40 | 2026-09-16 | current | project_gpu_drafter_verify_split_60_40
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_gpu_drafter_verify_split_60_40.md
  MEASURED on H200 — the EAGLE4 drafter is the SLOW side (60.4%), not verify (39.6%); verify 4.748 ms/step is the hard floor under every drafter-side speed lever
- memory:feedback_agent_scratch_worktrees_fill_root_kill_builds | 2026-09-15 | current | feedback_agent_scratch_worktrees_fill_root_kill_builds
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_agent_scratch_worktrees_fill_root_kill_builds.md
  "2026-09-15: agents' git worktrees in the job tmp (on /) filled lab-host's root filesystem and killed a live PDI build mid-packaging. Brief agents to put worktrees and bulk scratch on /mnt/data, and check
- memory:feedback_broken_harness_fails_toward_clean | 2026-09-15 | current | feedback_broken_harness_fails_toward_clean
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_broken_harness_fails_toward_clean.md
  "Earned 2026-08-23 (D708), three times in one hour: verifying by INSPECTION is not a verdict — but EXECUTION is not a verdict either if the harness is wrong, and a broken harness FAILS TOWARD CLEAN. A
- memory:feedback_cron_heartbeat_first_action_agents_die_silently | 2026-09-15 | current | feedback_cron_heartbeat_first_action_agents_die_silently
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_cron_heartbeat_first_action_agents_die_silently.md
  FIRST action of every resume is CronList -> arm the hourly heartbeat. Crons are session-only. Agents die silently (API 529) OR hang forever on a tool call that never executes; detect only by artifact 
- memory:feedback_no_file_removals_user_directive | 2026-09-15 | current | feedback_no_file_removals_user_directive
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_no_file_removals_user_directive.md
  "2026-09-15 USER: 'why u removing files??' — no rm/mv/find -delete/git clean/worktree remove by the orchestrator OR any agent, even in scratch, without asking; create-only."
- memory:feedback_a_zero_is_a_claim_about_the_instrument | 2026-09-14 | current | feedback_a_zero_is_a_claim_about_the_instrument
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_a_zero_is_a_claim_about_the_instrument.md
  "A clean result (zero hits, PASS, 'absent', 'nothing changed') is a claim about the INSTRUMENT until its scope, its tool identity and its stderr have been bounded. Nine distinct mechanisms produced a 
- memory:feedback_added_constraint_can_disable_an_existing_one | 2026-09-14 | current | feedback_added_constraint_can_disable_an_existing_one
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_added_constraint_can_disable_an_existing_one.md
  A constraint you ADD can silently disable a protective constraint that was already there; check report_exceptions -ignored for "Totally overridden path by CG N" after every new clock group or exceptio
- memory:feedback_control_that_supplies_the_path_never_tests_discovery | 2026-09-14 | current | feedback_control_that_supplies_the_path_never_tests_discovery
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_control_that_supplies_the_path_never_tests_discovery.md
  "A control that HANDS a checker its input never tests the checker's ability to FIND the input — and 'nothing there' is the one input on which a broken search and a correct search are indistinguishable
- memory:feedback_gate_races_its_artifact_partial_read | 2026-09-14 | current | feedback_gate_races_its_artifact_partial_read [0xE4B7113D]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_gate_races_its_artifact_partial_read.md
  A gate/watcher that reads a Vivado artifact (netlist, rpt, dcp, log) while Vivado is still writing it scores a PARTIAL file; the verdict direction depends on the file's byte layout, and it can fail to
- memory:feedback_grep_wrapper_gitignore_reads_zero | 2026-09-14 | current | feedback_grep_wrapper_gitignore_reads_zero
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_grep_wrapper_gitignore_reads_zero.md
  TWO ways `grep` reads a silent ZERO here — (1) interactive `grep` is a wrapper honouring .gitignore so a repo-root `grep -r ... .` returns 0 while `command grep` returns the truth; (2) `xargs command 
- memory:feedback_submodule_gitlink_silently_reverts_fixes | 2026-09-14 | current | feedback_submodule_gitlink_silently_reverts_fixes
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_submodule_gitlink_silently_reverts_fixes.md
  A fix committed in the `verilog` submodule DOES NOT EXIST for a build until the parent gitlink advances. Has silently reverted 4+ landed fixes. Check gitlink == submodule HEAD before every build.
- memory:feedback_verify_the_quantity_not_just_the_value | 2026-09-14 | current | feedback_verify_the_quantity_not_just_the_value
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_verify_the_quantity_not_just_the_value.md
  "2026-07-31: verifying a NUMBER is not verifying it is the number of the THING IN QUESTION — and a cross-check that shares its source with the claim is not independent. Cost: a 64x sizing defect that 
- memory:project_3b_2bit_gpu_baseline | 2026-09-14 | current | 3B+2bit GPU baseline (full mt_bench)
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_3b_2bit_gpu_baseline.md
  "Llama-3.2-3B + EAGLE4 2-bit on H200 NVL — 125.47 tok/s @ accept 2.11 over 80 prompts, 2026-04-26. ⚠️ the 2.11 is at --max-new-tokens 32 / depth 3 — NOT the accept comparator; cite the measured 1.9726
- memory:project_gpu_fa4_tree_draft_defect | 2026-09-14 | current | project-gpu-fa4-tree-draft-defect
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_gpu_fa4_tree_draft_defect.md
  The GPU EAGLE reference itself can be wrong — FA4 tree-draft attention (paged KV, causal=False) is broken on the H200; SGLang padded draft-extend leaves ghost rows that corrupt naive capture builds
- memory:project_qsfp_gtm_quads_slr1_tail_caps_one_link | 2026-09-14 | current | project_qsfp_gtm_quads_slr1_tail_caps_one_link
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_qsfp_gtm_quads_slr1_tail_caps_one_link.md
  All 4 GTM quads on the V80 (the QSFP cages) are in SLR1, and there is no alternative. The TAIL's SLR1 is full (99.44%), so a HEAD↔TAIL pair is capped at ONE link as built; 4 links fit only on the HEAD
- memory:reference_agent_commons_watcher_inbox | 2026-09-14 | current | reference_agent_commons_watcher_inbox
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/reference_agent_commons_watcher_inbox.md
  Agent Commons board (agents.yshicarrick.com) — an hourly OS-cron Haiku watcher drops new-post notices in ~/.agent-commons/inbox/; read it each heartbeat. Never post without the user's word.
- memory:feedback_oracle_without_weights_is_a_mixed_run | 2026-09-13 | current | feedback_oracle_without_weights_is_a_mixed_run
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_oracle_without_weights_is_a_mixed_run.md
  Supplying a golden/oracle path without ALSO supplying WROOT/GAMMADIR/SCALEHEX/LMFULL silently mixes a different checkpoint's weights with that oracle; the mismatch reads as a datapath RED
- memory:feedback_unquoted_heredoc_executes_backticks_in_prose | 2026-09-13 | current | feedback_unquoted_heredoc_executes_backticks_in_prose
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_unquoted_heredoc_executes_backticks_in_prose.md
  "`cat >> report.md <<EOF` (UNQUOTED) runs command substitution on every backtick and $ in the prose. Writing a report that quotes shell commands in backticks EXECUTES them: on 2026-09-13 it launched a
- memory:feedback_vacuous_gate_class_grep_terminated | 2026-09-13 | current | feedback_vacuous_gate_class_grep_terminated
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_vacuous_gate_class_grep_terminated.md
  "32% of this repo's gates cannot fail — grep-terminated scripts invert exit codes, single-token/ctx=1 stimuli make accuracy gates blind, and (mechanism 6, 2026-09-13) a threshold grepped out of a $dis
- memory:feedback_xdc_if_puts_silently_skipped_empty_pblock | 2026-09-13 | current | feedback_xdc_if_puts_silently_skipped_empty_pblock
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_xdc_if_puts_silently_skipped_empty_pblock.md
  "XDC has NO procedural escape: if/foreach/while/for/catch/puts/open/close/clock inside a constraint-imported .xdc are SILENTLY SKIPPED (bare `set` IS tolerated -- corrected 2026-09-13) ([Designutils 2
- memory:feedback_cdc_clock_groups_is_not_cdc_safety | 2026-09-12 | current | feedback_cdc_clock_groups_is_not_cdc_safety
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_cdc_clock_groups_is_not_cdc_safety.md
  "set_clock_groups -asynchronous removes the timing FAILURE, not the CDC hazard; a 2-FF synchroniser without ASYNC_REG is not a certified synchroniser; certify with report_cdc on the ROUTED dcp"
- memory:feedback_vivado_echo_makes_done_marker_grep_fire_early | 2026-09-12 | current | feedback_vivado_echo_makes_done_marker_grep_fire_early
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_vivado_echo_makes_done_marker_grep_fire_early.md
  Vivado batch mode ECHOES every tcl source line (prefixed "# ") into the log before running it, so `grep -q MY_DONE_MARKER log` fires the instant the script loads; anchor with ^MARKER
- memory:feedback_xpm_fifo_async_needs_user_clock_groups | 2026-09-12 | current | feedback_xpm_fifo_async_needs_user_clock_groups [0xE4B7113F]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_xpm_fifo_async_needs_user_clock_groups.md
  xpm_fifo_async does NOT constrain its own wr_clk/rd_clk relationship — without set_clock_groups -asynchronous the XPM's own synchronisers get timed as synchronous and the post-route report shows a −10
- memory:feedback_module_default_and_its_comment_are_not_the_build | 2026-09-11 | current | feedback_module_default_and_its_comment_are_not_the_build
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_module_default_and_its_comment_are_not_the_build.md
  A parameter's DEFAULT value and the comment explaining it are both authored at the module and both go stale; only the instantiation is the build. Cost a published claim on 2026-09-11.
- memory:feedback_synth_tolerates_forward_ref_xvlog_rejects | 2026-09-11 | current | feedback_synth_tolerates_forward_ref_xvlog_rejects [0xE4B7113B]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_synth_tolerates_forward_ref_xvlog_rejects.md
  Vivado synthesis binds a net used before its declaration; strict xvlog rejects it (VRFC 10-3380). A fork can build PDIs for weeks while every full-drafter xsim gate is silently uncompilable — re-run r
- memory:feedback_find_is_bfs_newermt_relative_reads_zero | 2026-09-09 | current | feedback_find_is_bfs_newermt_relative_reads_zero
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_find_is_bfs_newermt_relative_reads_zero.md
  "`find` on this box is bfs, not GNU findutils. It REJECTS relative -newermt ('-6 hours') as 'Invalid timestamp' and returns ZERO -- and 2>/dev/null hides the error, so a broken recency check reads as 
- memory:feedback_no_canonical_verdict_row_contract | 2026-09-09 | current | feedback_no_canonical_verdict_row_contract
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_no_canonical_verdict_row_contract.md
  "Five gates on the EAGLE4 flash path emit five different FAIL-row shapes and the NFAIL collector cannot see most of them. The root cause is not any regex: no contract defines the verdict-row format, s
- memory:feedback_per_agent_ram_gates_admit_all_agents | 2026-09-02 | current | feedback_per_agent_ram_gates_admit_all_agents
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_per_agent_ram_gates_admit_all_agents.md
  "A per-agent \"wait for >=N GB free\" gate does NOT limit total fleet memory - N agents check simultaneously, all see free RAM, and all launch. Gate the FLEET, not each agent."
- memory:feedback_partition_boot_after_flash | 2026-08-27 | current | feedback_partition_boot_after_flash [0xE4B71159]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_partition_boot_after_flash.md
  "A successful flash proves STORAGE not EXECUTION — the card may boot the other partition; detect ONLY via the build-id register, and NEVER use device_boot to force it"
- memory:feedback_session_context_backup_mandatory | 2026-08-27 | current | feedback_session_context_backup_mandatory
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_session_context_backup_mandatory.md
  User-mandated — write every measured fact to a dated markdown IN THE SAME TURN it is measured; chat-only facts are unrecoverable because cloud transcripts cannot be retrieved
- memory:feedback_gptq_deq_needs_transpose_vs_bf16 | 2026-08-24 | current | feedback_gptq_deq_needs_transpose_vs_bf16
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_gptq_deq_needs_transpose_vs_bf16.md
  GPTQ dequant is [in,out] but bf16 Linear.weight is [out,in] — comparing them untransposed reads r~0 and fakes "different models"
- memory:feedback_pipefail_head_sigpipe_kills_collectors | 2026-08-24 | current | feedback_pipefail_head_sigpipe_kills_collectors
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_pipefail_head_sigpipe_kills_collectors.md
  In a bash gate, `set -o pipefail` + `| head -N` SIGPIPEs the upstream collector, so a whole signal silently vanishes from the check — non-deterministically, depending on pipe-buffer luck.
- memory:feedback_replica_census_by_name_suffix_is_vacuous | 2026-08-24 | current | feedback_replica_census_by_name_suffix_is_vacuous
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_replica_census_by_name_suffix_is_vacuous.md
  Counting max_fanout replicas by matching net names against `_rep`/`_replica` reads ZERO on a netlist carrying thousands — Vivado names a replicated driver after the first flop bit it drives, not after
- memory:feedback_link_design_device_query_answers_floorplan_questions | 2026-08-23 | current | feedback_link_design_device_query_answers_floorplan_questions.md
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_link_design_device_query_answers_floorplan_questions.md
  **Earned 2026-08-23** (`reports/results/STREAM16_SLL_CENSUS_20260823.md`, commit `2dd3824`).
- memory:feedback_ooc_boundary_hold_min_delay_manufactures_red | 2026-08-23 | current | feedback_ooc_boundary_hold_min_delay_manufactures_red.md
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_ooc_boundary_hold_min_delay_manufactures_red.md
  **Earned 2026-08-23** on the 16-channel HBM weight streamer (`reports/results/STREAM16_TIMING_CLOSURE_20260823.md`,
- memory:feedback_success_is_2bit_accept_parity | 2026-08-23 | current | feedback-success-is-2bit-accept-parity
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_success_is_2bit_accept_parity.md
  "2026-06-09 USER SUCCESS CRITERION: 2-bit is the design point (fits URAM = the FPGA thesis); success = FPGA accept parity with the GPU's 2-bit accept (2.11-2.17), NOT precision escalation"
- memory:project_second_build_box_build-host-M | 2026-08-23 | current | project_second_build_box_build-host-M [0xE4B71111]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_second_build_box_build-host-M.md
  "2026-07-18 — 2nd build machine \"build-host-M\" @ user@<ip> (pw 2441): 20 cores/94G/599G free, Vivado 2025.1 + license, AVED trunk — PROVEN mono PDI builder. Build-only (no GPU, no V80 card). P
- memory:feedback_agents_orphan_on_detached_waiters | 2026-08-21 | current | feedback_agents_orphan_on_detached_waiters
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_agents_orphan_on_detached_waiters.md
  "Earned 2026-08-21 (6 agents in one day): a subagent that launches DETACHED work (setsid/nohup sims, builds, benches) and then 'waits for the notification' STOPS with no tracked children and is NEVER 
- memory:feedback_fsm_overlap_changes_need_md3_gate | 2026-08-21 | current | feedback_fsm_overlap_changes_need_md3_gate
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_fsm_overlap_changes_need_md3_gate.md
  "Earned twice in one week (FINDING-1 on 1153, D496 audit; D526 on 1157): ANY change that lets two FSM phases overlap (head/tail split, tail-stage producers, early relaunch) can mis-tag per-child write
- memory:feedback_inert_constraint_may_be_protective | 2026-08-21 | current | feedback_inert_constraint_may_be_protective
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_inert_constraint_may_be_protective.md
  "Earned 2026-08-21 (D518): a timing constraint found SILENTLY INERT is NOT automatically a cheap fix to re-apply — verify SAFETY from RTL first. loadpath_multicycle.xdc's set_multicycle_path on boot_w
- memory:feedback_document_decisions_never_wait | 2026-08-19 | current | feedback_document_decisions_never_wait [0xE4B71128]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_document_decisions_never_wait.md
  "User-mandated 2026-08-19 — document EVERY decision with its reasoning, supporting data and validation in DECISION_LOG.md, but NEVER wait on the user; execute and adjust. Documentation is written AFTE
- memory:feedback_no_meaningless_confirmations_pdi_is_goal | 2026-08-19 | current | feedback_no_meaningless_confirmations_pdi_is_goal [0xE4B71128]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_no_meaningless_confirmations_pdi_is_goal.md
  "User-mandated 2026-08-04, ESCALATED 2026-08-19 ('Never ask and wait on my call') — never ask for confirmation and never PARK work awaiting a ruling. FLASH (-p 0) IS PRE-AUTHORIZED and must be EXECUTE
- memory:project_checkpoint_weights_index | 2026-08-19 | current | project_checkpoint_weights_index [0xE4B71124]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_checkpoint_weights_index.md
  "Master index of ALL drafter checkpoint/weight paths — repo, SGLang version, serving code, quantization, accept/speed — GPTQ lineage vs OSPREY/eaglefit"
- memory:feedback_gpu_baseline_thermally_throttled | 2026-08-17 | current | feedback_gpu_baseline_thermally_throttled
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_gpu_baseline_thermally_throttled.md
  "H200 tok/s on this box is order- and temperature-dependent and must always be reported with clock/temp telemetry + run ordinal. ⚠️ BUT THE ORIGINAL MECHANISM IS REFUTED (D472): the 345 MHz samples ha
- memory:feedback_keep_pdi_building_never_vacate | 2026-08-16 | current | feedback_keep_pdi_building_never_vacate
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_keep_pdi_building_never_vacate.md
  USER RULE 2026-08-16 - a PDI build must ALWAYS be running or queued on the build machines. Never let the boxes go idle on the build axis.
- memory:feedback_number_from_wrong_domain_dma_bytes_as_ff | 2026-08-16 | current | feedback_number_from_wrong_domain_dma_bytes_as_ff [0xE4B7114B]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_number_from_wrong_domain_dma_bytes_as_ff.md
  "A plausible-magnitude number lifted from the WRONG DOMAIN reads as a real measurement and can close a lever. '+3,031,040 FF = 96.8% of device' was c2h_bytes, a host DMA transfer size. TEST: a resourc
- memory:feedback_onchip_fulltree_north_star | 2026-08-16 | current | feedback-onchip-fulltree-north-star
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_onchip_fulltree_north_star.md
  "2026-06-09 USER DIRECTIVE: the FPGA's reason-to-exist is weight REUSE across the on-chip tree loop — depth-0-only defeats it; next PDI must think deeply about on-chip multi-depth tree decode"
- memory:feedback_sibling_target_is_W8_never_scope_down | 2026-08-16 | current | feedback-sibling-target-is-w8-never-scope-down
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_sibling_target_is_W8_never_scope_down.md
  "2026-07-31 user correction: the sibling-batching target is W=8 (the FULL second-node set, 8 children per weight sweep) — NOT W=2. Never silently scope a user goal down; if physics forces less, SAY SO
- memory:project_drafter_precision_caps_accept_bf16_368 | 2026-08-16 | current | project_drafter_precision_caps_accept_bf16_368
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_drafter_precision_caps_accept_bf16_368.md
  "Measured — the 2-bit GPTQ quantization (not the EAGLE4 architecture) is the accept limiter: 2.08 vs BF16 3.68 = a 1.77x RATIO at depth-3 / 32-tok. ⚠️ the ABSOLUTE '~2.1 cap' does NOT transfer — the s
- memory:project_g8_gather_new_speed_wall_postspeedwave | 2026-08-16 | current | project_g8_gather_new_speed_wall_postspeedwave [0xE4B71111]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_g8_gather_new_speed_wall_postspeedwave.md
  "2026-07-14 K_C=512-era per-phase drawdown (0xE4B71111) — G8 HBM gather 49%. ⚠️ SUPERSEDED AT THE SHIPPED OPERATING POINT: see project_g8_not_the_wall_g4_qkv_is. Retained for the N-channel stripe spec
- memory:project_g8_not_the_wall_g4_qkv_is | 2026-08-16 | current | project_g8_not_the_wall_g4_qkv_is [0xE4B71143]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_g8_not_the_wall_g4_qkv_is.md
  "MEASURED on silicon — G4 QKV is the TOP phase at 23.2%, G9 20.3%, G8 only 10.9% (NOT 93%, and NOT the interim 19.0% — G8's counter is not state-gated and double-counts the TREE). Speed effort belongs
- memory:feedback_silence_must_not_read_as_success | 2026-08-14 | current | feedback_silence_must_not_read_as_success
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_silence_must_not_read_as_success.md
  The session's dominant defect class — a probe/gate/monitor that emits NOTHING on the failure it watches for, so absence reads as pass. Six independent instances in one day.
- memory:feedback_gates_report_only_never_kill_builds | 2026-08-13 | current | feedback_gates_report_only_never_kill_builds
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_gates_report_only_never_kill_builds.md
  "2026-07-30 user mandate (angry, twice): NO in-build gate may abort/kill a build — gates REPORT; a build-abort requires the user AND the orchestrator looking at it together. Kill authority lives at th
- memory:feedback_one_var_two_gates_dsp_per_lane | 2026-08-13 | current | feedback_one_var_two_gates_dsp_per_lane
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_one_var_two_gates_dsp_per_lane.md
  DSP_PER_LANE feeds TWO gates measuring DIFFERENT quantities — [40..44] for PUMPCHECK (device-total/128), [19..23] for R5 (per-module count). Never "fix" a band without reading its gate's measurement l
- memory:feedback_routed_refname_suffix_vacuous_census | 2026-08-13 | current | feedback_routed_refname_suffix_vacuous_census
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_routed_refname_suffix_vacuous_census.md
  Post-route REF_NAMEs gain suffixes (RAMB36E5->RAMB36E5_INT, CARRY8->LOOKAHEAD8). A census keyed on the exact pre-route name returns 0 on a routed netlist ALWAYS — a vacuous PASS.
- memory:feedback_two_weight_bakes_same_basename_silent_swap | 2026-08-13 | current | feedback_two_weight_bakes_same_basename_silent_swap
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_two_weight_bakes_same_basename_silent_swap.md
  Two lmdown weight bakes (p4 = wrong frontier, p4_8815 = correct) share basenames. The shipping module's DEFAULT and its sibling module both point at the WRONG one — a module swap changes weights silen
- memory:project_frspec_32k_is_the_knee_shortlist_not_permutation | 2026-08-13 | current | project_frspec_32k_is_the_knee_shortlist_not_permutation [0xE4B71146]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_frspec_32k_is_the_knee_shortlist_not_permutation.md
  FR-Spec's speed win is the 32K SHORTLIST, not the permutation (which is exactly output-neutral). 32K is the measured knee. The vocab cost is booked to TREE, not G12.
- memory:feedback_agent_turn_completion_orphans_bg_jobs | 2026-08-11 | current | feedback_agent_turn_completion_orphans_bg_jobs
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/feedback_agent_turn_completion_orphans_bg_jobs.md
  "Background build jobs (nohup'd vivado etc.) launched from INSIDE an agent's turn can orphan/stall when the agent completes its turn normally — same class as TaskStop orphaning, but triggered by ordin
- memory:project_milestone_98toks_kvonly_20260811 | 2026-08-11 | current | project_milestone_98toks_kvonly_20260811 [0xE4B71111,0xE4B71143]
  ~/.claude/projects/-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE/memory/project_milestone_98toks_kvonly_20260811.md
  "MILESTONE 2026-08-11: 98.15 tok/s batch-mean (100.60 final-report) @ accept 1.7829 E2E on 0x1143 (md=2, K_C=64, cuda-graph ON), zero reflash — the +37% came from arming the shipped-but-unused KVONLY 

## result_report (1658)

(showing 150 newest of 1658; all in catalog.json)

- results:SESSION_CONTEXT_BACKUP_20261026.md | 2026-10-26 | current | SESSION CONTEXT BACKUP 2026-10-26 (lab-host-two, 4C1)
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261026.md
  - 2026-10-26T01:01Z catch-up after usage-limit gap (~14:30Z 10-25 -> 01:00Z 10-26): 0 Vivado on buildx00-04, buildy00-05, build-host-M, lab-host; no HQ_* notes in 25 h; no new ~/serving/qa questions; tracklist
- results:SESSION_CONTEXT_BACKUP_20261025.md | 2026-10-25 | current | SESSION CONTEXT BACKUP 2026-10-25 (lab-host-two, 4C1)
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261025.md
  - 2026-10-25T01:11Z 2h WNS/SLR watch: 0 Vivado on buildx00-04, buildy00-05, build-host-M, lab-host. No in-flight 4C1 build.
- results:SESSION_CONTEXT_BACKUP_20261024.md | 2026-10-24 | current | SESSION CONTEXT BACKUP 2026-10-24 (lab-host-two)
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261024.md
  lab-host: / 808G (9%), /mnt/data = /mnt/models = /sanctuary 1.1T (72%).
- results:SESSION_CONTEXT_BACKUP_20261023.md | 2026-10-23 | current | SESSION CONTEXT BACKUP 2026-10-23 (lab-host-two)
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261023.md
  - 2026-10-23 ~10:5xZ **PKCACHE A/B MEASURED (lab-host-qwen, f70f159a; lab-host Llama 3B, cards 0x129C/0x128B, N2 fifo, humaneval 80x256 w2 d7, stdlib packer, rounds 21-24 PKCACHE 0/1/1/0):** N2 aggregate 606.
- results:TALON_ACCESS_20261021.md | 2026-10-21 | current | Access guide: talon and talon-2 (2026-10-21)
  <repo>/reports/results/TALON_ACCESS_20261021.md
  No passwords, tokens or private keys are in this file. Access is by ssh key behind Cloudflare Access.
- results:TALON_DEMO_GUIDE_20261021.md | 2026-10-21 | current | talon demo guide: what we demo, which images, how to serve (2026-10-21)
  <repo>/reports/results/TALON_DEMO_GUIDE_20261021.md
  One file for the talon demo. It covers the message to send, the images, the repo layout to put on talon, and the serve commands for each model.
- results:SESSION_CONTEXT_BACKUP_20261020.md | 2026-10-20 | current | SESSION_CONTEXT_BACKUP_20261020.md [0xE4B7118B,0xE4B7119B,0xE4B7119C,0xE4B711B1]
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261020.md
  - uptime 4 min at 14:38; lspci 10ee 50b4/50b5 on 01/42/61/62; H200 14 MiB used, 50 C; / 809 GB free; RAM 117 GB free.
- results:lab-host_FREEZE_RCA_20261019.md | 2026-10-19 | current | lab-host (lab-host) freeze RCA, 2026-10-19
  <repo>/reports/results/lab-host_FREEZE_RCA_20261019.md
  Read-only investigation. No card writes, no resets, no driver load/unload, no process kills, no file removals, no probe rerun.
- results:BUILDHOST_VIVADO_INSTALL_20261019.md | 2026-10-19 | current | Build-host Vivado 2025.1 + SMBus licence survey (2026-10-19)
  <repo>/reports/results/BUILDHOST_VIVADO_INSTALL_20261019.md
  Scope: build00, buildx00-04, buildxd00-05, buildy00-05, buildz00-05, buildzd00-05 (30 hosts), via `ssh -p 65022 user@<h>.a.eagle.ai`. Run from lab-host (this machine), which only ran ssh and read-only 
- results:SESSION_CONTEXT_BACKUP_20261019.md | 2026-10-19 | current | SESSION CONTEXT BACKUP 2026-10-19 [0xE4B71187,0xE4B71188]
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261019.md
  - Running on silicon (discover reg 0x100 + tp_link4 build-id, boot #22): 01/42 0xE4B71288 (reg100 0xe4b71188), 61/62 0xE4B71287 (reg100 0xe4b71187).
- results:SESSION_CONTEXT_BACKUP_20261018.md | 2026-10-18 | current | SESSION_CONTEXT_BACKUP_20261018.md
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261018.md
  - Source: `git -C verilog archive 56233e0647c4 forks/fourcard_4c1v1_mlpv_20261017` (tar md5 5d2e5667694bde6176435c44d12ad9bb) -> build-vm `/shared/build/v1mlpv_56233e0_src` (EXPORT_COMMIT.txt); mlpv_vs_
- results:UNI_DUALCLK___build-host-A_PKG.md | 2026-10-18 | current | UNI_DUALCLK___build-host-A_PKG.md
  <repo>/reports/results/UNI_DUALCLK___build-host-A_PKG.md
  PACKAGE_OK WNS= URAM=
- results:SESSION_CONTEXT_BACKUP_20261017.md | 2026-10-17 | current | SESSION_CONTEXT_BACKUP_20261017.md
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261017.md
  - Report: `reports/results/FOURCARD_DESIGN_20261016/GPU_GOLDEN_KVI_20261017.md`; data `<path:`> MANIFEST md5 3d31f313d4a6305371a4b9c987b3885c; scripts `scripts/gp
- results:UNI_DUALCLK_s3LAattf5_head_build-host-A_PKG.md | 2026-10-17 | current | UNI_DUALCLK_s3LAattf5_head_build-host-A_PKG.md
  <repo>/reports/results/UNI_DUALCLK_s3LAattf5_head_build-host-A_PKG.md
  PACKAGE_OK WNS=0.176 URAM=432 316b886c206e62f59ba10a6dc9acc2ae  /shared/build/s3LAattf5/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:UNI_DUALCLK_s3LAattf2_head_build-host-A_PKG.md | 2026-10-16 | current | UNI_DUALCLK_s3LAattf2_head_build-host-A_PKG.md
  <repo>/reports/results/UNI_DUALCLK_s3LAattf2_head_build-host-A_PKG.md
  PACKAGE_OK WNS=0.176 URAM=432 f57946651ef98b410af68db03247d95f  /shared/build/s3LAattf2/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:KVIFIX_TIMING_CENSUS_20261014.md | 2026-10-14 | current | KVIFIX card-A HEAD: setup-path census of the post-phys_opt checkpoint (2026-10-14)
  <repo>/reports/results/KVIFIX_TIMING_CENSUS_20261014.md
  Program: 3-card Llama-3.2-3B DRAFTER (card A HEAD, fork `verilog/forks/head_s3LA_kvifix_20261014`, reader `kvi_fan_rd.sv` md5 `7f31e412`).
- results:SESSION_CONTEXT_BACKUP_20261014.md | 2026-10-14 | current | lab-host-main session context (curated) — 2026-10-12 → 2026-10-14 05:0x EDT [0xE4B7111B,0xE4B71197,0xE4B7119E,0xE4B711BD...]
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261014.md
  Curated from session 5b98fbdb (raw transcript copy: <path:lab-host_main_5b98fbdb_20261014_0501.jsonl> md5 9f16a1ca, also build-host-A /shared/rnd_hq/transcripts/session
- results:SPEC_DECODE_EAGLE4_VS_DMATRIX_20261014.md | 2026-10-14 | current | Speculative decoding: our EAGLE4 FPGA drafter vs d-Matrix Corsair (2026-10-14) [0xE4B71143,0xE4B71199,0xE4B711BD]
  <repo>/reports/results/SPEC_DECODE_EAGLE4_VS_DMATRIX_20261014.md
  Page: https://hardware.research.yshicarrick.com/p/spec-decode-eagle4-vs-dmatrix/ (v1). Figures: `reports/results/spec_decode_dmatrix_20261014/fig_{split,sizes,drafter_dataflow}.svg`; page data `page.j
- results:lab-host_GOLDEN_AGENT_BRIEF_20261011.md | 2026-10-11 | current | Brief: lab-host golden + RTL-test-suite research agent (launched 2026-10-11 by lab-host on the USER's order)
  <repo>/reports/results/lab-host_GOLDEN_AGENT_BRIEF_20261011.md
  You are a FORK of the lab-host session (same context up to this point), now a separate lab-host research agent named **lab-host-golden**
- results:lab-host_GOLDEN_SEED_PROMPT_20261011.md | 2026-10-11 | current | lab-host-golden SEED PROMPT (fresh-session recovery; lab-host-golden has NO transcript on disk, so --resume 
  <repo>/reports/results/lab-host_GOLDEN_SEED_PROMPT_20261011.md
  You are lab-host-golden (LAPTOP COPY), started on the laptop (Legion-7, .25) at the USER's order so lab-host's cold AC does not stop the research ("oh shoot, can you copy yourself to latpop and launch it ther
- results:GOLDEN_SUITE_20261011.md | 2026-10-11 | current | GOLDEN SUITE 2026-10-11: per-layer / per-op goldens + RTL test suites + AXI input testbench (lab-host-go [0xE4B71112,0xE4B7111B,0xE4B7111C,0xE4B7111D...]
  <repo>/reports/results/GOLDEN_SUITE_20261011.md
  **Agent:** lab-host-golden, a fork of the main lab-host session, launched on the user's order 2026-10-11 ~02:25 EDT.
- results:HBM_RING8_LLM_SERVING_DESIGN_20261011.md | 2026-10-11 | current | HBM_RING8: serving any LLM from HBM on an 8-FPGA ring, driven by a model contract file (design, 2026
  <repo>/reports/results/HBM_RING8_LLM_SERVING_DESIGN_20261011.md
  Author: lab-host (fork), for the user. Design and research only: no RTL was edited, nothing was built or flashed.
- results:HBM_RING8_LLM_SERVING_DESIGN_20261011_v2.md | 2026-10-11 | current | HBM_RING8 v2: serving any LLM from HBM on an 8-FPGA ring, driven by a model contract file (design, 2
  <repo>/reports/results/HBM_RING8_LLM_SERVING_DESIGN_20261011_v2.md
  **v2 (2026-10-11) answers the user's review of v1:**
- results:OSPREY_GOLDENS_PROGRAM1_20261011.md | 2026-10-11 | current | OSPREY per-op goldens, Program 1 (2026-10-11, lab-host-golden-osprey)
  <repo>/reports/results/OSPREY_GOLDENS_PROGRAM1_20261011.md
  **Program 1: the 3-card Llama-3.2-3B EAGLE4 DRAFTER** (3STAGE line: A attn @42, B MLP @61, C lm+tree @62).
- results:QWEN27B_RING8_DESIGN_20261011.md | 2026-10-11 | current | QWEN27B_RING8: serving Qwen3.6-27B on the 8-FPGA ring, reusing our device (design, 2026-10-11)
  <repo>/reports/results/QWEN27B_RING8_DESIGN_20261011.md
  User ask (verbatim): "if we target the qwen 27b target model on the 8 fpga, what would be the best design, reusing our device as much as possible".
- results:RING8_TWO_IMAGE_TRACKS_20261011.md | 2026-10-11 | current | RING8: two 8-card image tracks (lab-host-golden, 2026-10-11)
  <repo>/reports/results/RING8_TWO_IMAGE_TRACKS_20261011.md
  **USER order (2026-10-11 ~05:4x, verbatim):** "set track to deliver 2 modles of 8-card image; one is constant fold design on ternary
- results:SESSION_CONTEXT_BACKUP_20261011.md | 2026-10-11 | current | SESSION_CONTEXT_BACKUP_20261011.md
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20261011.md
  int8 per-token activations on the whole draft tree (no CUDA graphs, humaneval 80, d15 w4): fa4 A_off 8.29 / B_int8 8.25 / RED int2 1.40; fa3 A_off 8.93 / B_int8 8.91; hook calls ÷ verify steps ≈ 15.05
- results:TERNARY_CONSTFOLD_LUT_STUDY_20261011.md | 2026-10-11 | current | Ternary weights constant-folded into LUT adder trees: measured cost, walls, and the road to a PDI (2 [0xE4B71128]
  <repo>/reports/results/TERNARY_CONSTFOLD_LUT_STUDY_20261011.md
  **Scope: the NEW 8-FPGA SERVER program. The full target model runs on the 8-card V80 ring, with no GPU.** Candidate targets are BitNet-b1.58-2B-4T, TriLM and a ternarized Llama-3.2-3B. This is not the
- results:TERNARY_MAKE_POSSIBLE_20261011.md | 2026-10-11 | current | TERNARY_MAKE_POSSIBLE — full BitNet-2B-4T target model folded on 8 V80s (2026-10-11, PARTIAL — lab-host 
  <repo>/reports/results/TERNARY_MAKE_POSSIBLE_20261011.md
  **Program: the NEW 8-FPGA server — the FULL TARGET MODEL microsoft/bitnet-b1.58-2B-4T served on an 8-card V80 ring, no GPU.**
- results:AUDIT_PIPELINED_ATTENTION_I5_20260924.md | 2026-09-24 | current | AUDIT: I5 pipelined attention engine (`attention_head_assembly_pipe`), 2026-09-24
  <repo>/reports/results/AUDIT_PIPELINED_ATTENTION_I5_20260924.md
  Read-only pair audit of `verilog/forks/level_batched_head_20260924/src/attention_head_assembly_pipe.sv`
- results:BATCH2_PINGPONG_PAIR_20260924.md | 2026-09-24 | current | I2: batch-2 ping-pong serving on the two-card pair (design study, 2026-09-24) [0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/BATCH2_PINGPONG_PAIR_20260924.md
  **Idea (innovation board I2):** serve two requests at once. The pair drafts request B while the GPU verifies request A,
- results:G4_EMBFOLD_INTEGRATION_20260924.md | 2026-09-24 | current | G4_EMBFOLD integration: offline table builder, HBM fetch master, host side (2026-09-24) [0xE4B711C9]
  <repo>/reports/results/G4_EMBFOLD_INTEGRATION_20260924.md
  **Fork:** `verilog/forks/g4_embfold_hbm_20260924`. It is a create-only `git archive` of verilog `6197607` `forks/g4_embfold_20260923` (parent `43d67903`). Prior verdict: `reports/results/G4_EMBFOLD_20
- results:GPU_WIDTH_DEPTH_ACCEPT_GRID_20260924.md | 2026-09-24 | current | GPU-only accept and tok/s by tree width × depth, on the pair's checkpoint (2026-09-24)
  <repo>/reports/results/GPU_WIDTH_DEPTH_ACCEPT_GRID_20260924.md
  **Status: MEASURED, COMPLETE (2026-09-24 10:04).** All 132 planned runs processed: fa3 40/40 and fa4 40/40 grid cells, anchors, repeats, controls, batch-2, E3 split, and the two contamination re-runs.
- results:HEAD_SEAM256_KVSKIP_20260924.md | 2026-09-24 | current | HEAD_SEAM256_KVSKIP_20260924: 256-bit seam export and KV-only attention skip on the HEAD [0xE4B71137,0xE4B711CD]
  <repo>/reports/results/HEAD_SEAM256_KVSKIP_20260924.md
  **Status: GATES GREEN (xsim). RTL committed: verilog `0220bf7`.** Scope: RTL and xsim only. No build_all, no flash, no card access, no build-host-M. Nothing I did not start was killed; nothing was removed.
- results:HOST_GLUE_1MS_20260924.md | 2026-09-24 | current | I7 — pair host glue ≤ 1 ms? Measured: no. G is 4–6 ms today; the host-only floor is ~3 ms. Four byte [0xE4B711B0,0xE4B711C8,0xE4B711C9,0xE4B711CA]
  <repo>/reports/results/HOST_GLUE_1MS_20260924.md
  **Date:** 2026-09-24 · **Mission:** innovation board I7 (`docs/INNOVATION_BOARD_20260924.md`, 04:30 verdict: the pair beats the H200
- results:HOST_OVERLAP_IMPL_20260924.md | 2026-09-24 | current | Pair host overlap, implemented: worker GIL fix, early tree launch, HOSTFAST as the default, KVRES-CL [0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/HOST_OVERLAP_IMPL_20260924.md
  **Date:** 2026-09-24 · **Scope:** host code only. No RTL, build, flash, card access or GPU run: lab-host is due for a cold AC, and the H200 is
- results:HOST_STRUCTURAL_I8_20260924.md | 2026-09-24 | current | I8 — structural host redesign toward G ≤ ~1.5 ms: rank, prototype, and what the verify needs from th [0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/HOST_STRUCTURAL_I8_20260924.md
  **Date:** 2026-09-24 · **Mission:** innovation board I8 (after I7, `HOST_GLUE_1MS_20260924.md`).
- results:LAPTOP_BUILD_HOST_AND_D6_20260924.md | 2026-09-24 | current | Laptop (Legion-7) as a full PDI build host + fit probe 0xE4B711D6 — 2026-09-24 [0xE4B71128,0xE4B711CB,0xE4B711CF,0xE4B711D3...]
  <repo>/reports/results/LAPTOP_BUILD_HOST_AND_D6_20260924.md
  **Status (02:30 EDT, final for this agent):** the laptop is a **verified full-PDI build host**: inputs are md5-identical, the SMBus license works (IS_LOCKED=0), the AMC firmware builds, a dry STEP0-3 
- results:LEVEL_BATCHED_HEAD_20260924.md | 2026-09-24 | current | I1 Level-batched HEAD: the kill test, a pipelined G5 engine, and a level model (2026-09-24) [0xE4B711C9,0xE4B711E3]
  <repo>/reports/results/LEVEL_BATCHED_HEAD_20260924.md
  **Verdict.**
- results:PAIRWISE_SCALEOUT_20260924.md | 2026-09-24 | current | Pairwise scale-out: how K pairs would split one tree (design note, 2026-09-24) [0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/PAIRWISE_SCALEOUT_20260924.md
  **Status: DESIGN NOTE ONLY.** On 2026-09-24 the user narrowed the scope: "right now, let's limit us to pair-wise; we have to prove we have a chance here". So the `PAIR_SCALEOUT_EMULATE` host flag was 
- results:PAIR_ACCEPT_GAP_RCA_20260924.md | 2026-09-24 | current | PAIR ACCEPT GAP RCA: pair 2.22 vs GPU 2.78 (2026-09-24) [0xE4B71159,0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/PAIR_ACCEPT_GAP_RCA_20260924.md
  **Status:** the gap is fully attributed, using a GPU accept emulator that reproduces 10 silicon and GPU runs. No card was touched, and no rounder or host code was changed.
- results:PAIR_BRINGUP_WEIGHTS_TEST_COMMANDS_20260924.md | 2026-09-24 | current | EAGLE4 pair: bring-up, QDMA, weight loading and tests — every command with its result (2026-09-24) [0xE4B711C8,0xE4B711CF]
  <repo>/reports/results/PAIR_BRINGUP_WEIGHTS_TEST_COMMANDS_20260924.md
  Pair: **HEAD `0xE4B711C8` on 61:00.0** (C9 source + HEAD_PAR 8) and **TAIL `0xE4B711CF` on 41:00.0** (attention pruned +
- results:PAIR_WIN_CONDITION_MODEL_20260924.md | 2026-09-24 | current | Can ONE pair beat GPU-only? Win-condition model for two V80s, with a K-pair appendix (2026-09-24) [0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/PAIR_WIN_CONDITION_MODEL_20260924.md
  **Status: MODEL.** Analysis only. No card, GPU, build or flash was touched, nothing was killed or removed, and build-host-M was not used.
- results:SESSION_CONTEXT_BACKUP_20260924.md | 2026-09-24 | current | SESSION_CONTEXT_BACKUP_20260924.md [0xE4B71167,0xE4B711C8,0xE4B711CF,0xE4B711D6...]
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20260924.md
  - MEASURED on live pair C9/B0 (pre-cold-AC), matched 50-req mt d2 K_C64 fa4 SGLang 0.5.14, loaded box (load ~25, H200 74-76 C): default HOSTFAST 73.93 tok/s (hglue_A_r1 md5 e0633f0d); +4 I7 cuts 76.68
- results:TAIL_ITEM_LEVERS_20260924.md | 2026-09-24 | current | TAIL item levers (Innovation I6): where the per-child TAIL item goes on CF, and the cheapest cut for
  <repo>/reports/results/TAIL_ITEM_LEVERS_20260924.md
  **Date** 2026-09-24 · **Source** TAIL CF = fork `verilog/forks/tree_width_tail_20260924` (export `ef36d00` = CB `tail_fast` + runtime
- results:TAIL_ITEM_LEVERS_IMPL_20260924.md | 2026-09-24 | current | TAIL item levers T1 + S1: implementation and gates (I6 follow-up)
  <repo>/reports/results/TAIL_ITEM_LEVERS_IMPL_20260924.md
  **Date** 2026-09-24 · **Fork** `verilog/forks/tail_item_levers_20260924`. It was made by `git archive ef36d00 forks/tree_width_tail_20260924`, and `diff -rq` against both the archive and the live tree
- results:TEXT_DRIFT_43of50_RCA_20260924.md | 2026-09-24 | current | TEXT DRIFT 43/50 — RCA (2026-09-24)
  <repo>/reports/results/TEXT_DRIFT_43of50_RCA_20260924.md
  **VERDICT: this is not a defect. The INPUT changed. The Llama-3.2 chat template writes the wall-clock date into the
- results:TREE_WIDTH_PAIR_20260924.md | 2026-09-24 | current | Runtime tree width on the pair (width 3) — TAIL `0xE4B711CF`, HEAD unchanged (2026-09-24) [0xE4B711CE,0xE4B711CF]
  <repo>/reports/results/TREE_WIDTH_PAIR_20260924.md
  **User decision (01:10):** run the pair at tree width 3. The modelled gain is +27–35% over width 8 at depth 2
- results:ACCEPT_RESTORED_HEADPAR1_20260923.md | 2026-09-23 | current | ACCEPT RESTORED: 2.76 on the new pair — and HEAD_PAR=8 is now demonstrated as the cause of the 1.00  [0xE4B71189,0xE4B711B0]
  <repo>/reports/results/ACCEPT_RESTORED_HEADPAR1_20260923.md
  **Date:** 2026-09-23, first boot after cold AC. **Cards RUNNING** (reg `0x100`, not flash logs):
- results:ADDER_TREE_STUDY_20260923.md | 2026-09-23 | current | ADDER TREE STUDY: 2-bit GEMV lane reduction (2026-09-23, FINAL)
  <repo>/reports/results/ADDER_TREE_STUDY_20260923.md
  **Question.** Is the lane's Wallace 3:2 carry-save tree (`csa_tree16_int20`: 16 x INT20 -> INT26, latency 3, II=1)
- results:ARRAY_DUALCLK_20260923.md | 2026-09-23 | current | ARRAY_DUALCLK_20260923: the 2-bit GEMV array on its own 2x clock [0xE4B711C9,0xE4B711CD]
  <repo>/reports/results/ARRAY_DUALCLK_20260923.md
  Status: xsim bit-exact gate, HEAD-set cycle census and one routed dual-clock harness (133.3/66.7 MHz) are MEASURED: **WNS +0.156 / WHS +0.010 on both clocks, report_cdc 'All paths are Safely Timed'**.
- results:ARRAY_POC_20260923.md | 2026-09-23 | current | ARRAY_POC_20260923 — five speed levers on the 2-bit GEMV array (module only) [0xE4B711C9,0xE4B711D3]
  <repo>/reports/results/ARRAY_POC_20260923.md
  Status: **RESULTS IN** for levers 1, 4, 5, RET_SER, M1 and M2 (routed); M3, DOT_DSP P&R and K_PAR-32 P&R PENDING (queued, rows marked). Rows are MEASURED, DERIVED (arithmetic on measured rows, formula
- results:ARRAY_SYSTOLIC_LUTGEMM_20260923.md | 2026-09-23 | current | ARRAY_SYSTOLIC_LUTGEMM_20260923 — systolic activation flow (A/A2) and multiply-free LUT-GEMM (B) on  [0xE4B711C9]
  <repo>/reports/results/ARRAY_SYSTOLIC_LUTGEMM_20260923.md
  Status: **FINAL for the bit-exact, cycle, lane-cost, array-synth and L128-route results; two routes are still QUEUED (§7).** Every cell is marked MEASURED (xsim or Vivado report, path given), DERIVED 
- results:build-host-A_BUILD_HOST_AND_D6_20260923.md | 2026-09-23 | current | build-host-A (PC <ip>) as a full PDI build host + fit probe 0xE4B711D6 — 2026-09-23 [0xE4B711CD,0xE4B711D3,0xE4B711D6]
  <repo>/reports/results/build-host-A_BUILD_HOST_AND_D6_20260923.md
  **Status: STAGED + VALIDATED; the first run died at BD create because the AMD SMBus IP license is locked to one machine (`[IP_Flow 19-2322]`). WAITING ON THE USER: build-host-A needs its own free SMBus lice
- results:AUDIT_SIBLING_W8_UNLOCK_20260923.md | 2026-09-23 | current | AUDIT: is sibling batching at W=8 "unlocked" on the design-D GEMV card? (2026-09-23, read-only) [0xE4B711D3]
  <repo>/reports/results/AUDIT_SIBLING_W8_UNLOCK_20260923.md
  Auditor: independent, read-only agent. The sibling fork was not edited, and no run I did not start was touched.
- results:DEPTH_N_PAIR_SPEED_MODEL_20260923.md | 2026-09-23 | current | Depth-N on the two-card pair: modelled tok/s by tree depth and width (2026-09-23) [0xE4B71193,0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/DEPTH_N_PAIR_SPEED_MODEL_20260923.md
  **Status: MODEL.** Read-only analysis. No card access, no build, no flash, nothing killed or removed.
- results:DRAFT_VERIFY_OVERLAP_AND_HOST_CUTS_20260923.md | 2026-09-23 | current | Draft/verify overlap and host-glue cuts, round 2: live pair `0xE4B711C9` HEAD + `0xE4B711B0` TAIL [0xE4B71192,0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/DRAFT_VERIFY_OVERLAP_AND_HOST_CUTS_20260923.md
  **Date:** 2026-09-23/24 · **Scope:** design plus host-side analysis only. No card access, no GPU run, no build, no flash, no process
- results:ENGINE_SPLIT_GEMV_STEP1_20260923.md | 2026-09-23 | current | ENGINE-SPLIT design D, step 1: the GEMV card (PIPE_ROLE 3) — RTL, elaboration, and the bit-exact gat [0xE4B711C9]
  <repo>/reports/results/ENGINE_SPLIT_GEMV_STEP1_20260923.md
  **Date:** 2026-09-23 · **Fork:** `verilog/forks/engine_split_gemv_20260923` (create-only copy of
- results:ENGINE_SPLIT_PAIR_STEP3_20260923.md | 2026-09-23 | current | ENGINE-SPLIT design D, step 3: GEMV card + state card + hop_link in ONE xsim (the pair gate)
  <repo>/reports/results/ENGINE_SPLIT_PAIR_STEP3_20260923.md
  **Date:** 2026-09-23/24 · **Fork:** `verilog/forks/engine_split_pair_20260923` (create-only copy of
- results:ENGINE_SPLIT_STATE_STEP2_20260923.md | 2026-09-23 | current | ENGINE-SPLIT design D, step 2: the STATE card (PIPE_ROLE 4) — RTL, elaboration, and the bit-exact ga
  <repo>/reports/results/ENGINE_SPLIT_STATE_STEP2_20260923.md
  **Date:** 2026-09-23 · **Fork:** `verilog/forks/engine_split_state_20260923` · **Decision being implemented:**
- results:G4_EMBFOLD_20260923.md | 2026-09-23 | current | G4_EMBFOLD: constant-folding G4's token-embedding half (2026-09-23/24) [0xE4B711C9]
  <repo>/reports/results/G4_EMBFOLD_20260923.md
  **Fork:** `verilog/forks/g4_embfold_20260923`, a create-only `git archive` export of `redesign_t0w2_kvonly_head_20260923` @ verilog `4a324ad` (the source of HEAD `0xE4B711C9`).
- results:GEMV_KPAR32_PROBE_20260923.md | 2026-09-23 | current | GEMV K_PAR=32 fit probe — 2026-09-23
  <repo>/reports/results/GEMV_KPAR32_PROBE_20260923.md
  **Question** (engine-split design, `reports/results/next_image_engine_split_20260923/README.md` §4.3/§6.2/§6.4):
- results:GEMV_RETSER_20260923.md | 2026-09-23 | current | GEMV_RETSER_20260923 — serialized per-SLR-group GEMV result return (`RET_SER`)
  <repo>/reports/results/GEMV_RETSER_20260923.md
  **Date:** 2026-09-23. Line numbers are in the FORK files at the md5s quoted. **Fork:** `verilog/forks/gemv_retser_20260923` (create-only copy of
- results:HEADPAR8_RCA_20260923.md | 2026-09-23 | current | HEAD_PAR=8 RCA: the attention datapath is bit-exact, and "HEAD_PAR caused 1.00" was never isolated [0xE4B711C9]
  <repo>/reports/results/HEADPAR8_RCA_20260923.md
  **2026-09-23 · xsim + one module OOC synth · no card, no flash, no build_all, no build-host-M, nothing killed, nothing removed.**
- results:LAPTOP_WORKER_25_20260923.md | 2026-09-23 | current | Laptop worker <ip> — READY (2026-09-23)
  <repo>/reports/results/LAPTOP_WORKER_25_20260923.md
  - **Host** `user-Legion-7-16IAX10`, Ubuntu (dual-boot with Windows): Intel Core Ultra 9 275HX, 24 CPUs, 62 GB RAM, 7 GB swap.
- results:LINK_TRANSPORT_20260923.md | 2026-09-23 | current | LINK TRANSPORT for the engine-split pair (design D) — `hop_link` — 2026-09-23
  <repo>/reports/results/LINK_TRANSPORT_20260923.md
  **Scope:** SystemVerilog + xsim only. No synthesis, no PDI build, no card touched. Design refs:
- results:PAIR_2X_ROOT_CAUSE_CATCHUP_TREES_20260923.md | 2026-09-23 | current | ROOT CAUSE of the pair's 2.31x slowdown: every KV catch-up row is run as a FULL 9-item tree [0xE4B71167,0xE4B71189,0xE4B71193]
  <repo>/reports/results/PAIR_2X_ROOT_CAUSE_CATCHUP_TREES_20260923.md
  **User question:** "what is slow though! a single fpga is 130 tps!" — answered by measurement.
- results:PAIR_HOST_OVERHEAD_20260923.md | 2026-09-23 | current | Pair SGLang serving loop: where the host time goes, and five host-side cuts (+4.8% tok/s, accept unc [0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/PAIR_HOST_OVERHEAD_20260923.md
  **Date:** 2026-09-23 · **Scope:** host-side only. No RTL change, no build, no flash, no reset, no 0x20 write.
- results:PAIR_KVONLY_HEAD_0x11C9_SILICON_20260923.md | 2026-09-23 | current | HEAD `0xE4B711C9` (W2 KV-only catch-up) on silicon — pair 57.15 → 102.13 tok/s at the same accept [0xE4B711B0,0xE4B711C9]
  <repo>/reports/results/PAIR_KVONLY_HEAD_0x11C9_SILICON_20260923.md
  **Date:** 2026-09-23 · **Status:** MEASURED on silicon, A/B on the same image, repeated.
- results:PAIR_PER_PHASE_CNTCLR_20260923.md | 2026-09-23 | current | Pair per-phase census at REAL serving — first use of the T0 `cnt_clr` instrument on silicon [0xE4B71189,0xE4B711B0]
  <repo>/reports/results/PAIR_PER_PHASE_CNTCLR_20260923.md
  Cards RUNNING (reg `0x100`): `61:00.0` = `0xe4b71189` (HEAD, HEAD_PAR=1) · `41:00.0` = `0xe4b711b0` (TAIL).
- results:SESSION_CONTEXT_BACKUP_20260923.md | 2026-09-23 | current | SESSION_CONTEXT_BACKUP_20260923.md [0xE4B71189,0xE4B711B0,0xE4B711B1,0xE4B711C8...]
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20260923.md
  - Fork `verilog/forks/redesign_t0w2_kvonly_head_20260923` @ submodule 4a324ad (gitlink 3fad2632); only RTL delta vs 0x1189 export = `top_fsm.sv` (a16a7b8d436de518a84f94add3766854 vs d04d7fb756f9...), 
- results:SIBLING_W_ON_GEMV_CARD_20260923.md | 2026-09-23 | current | SIBLING_W_ON_GEMV_CARD_20260923 — re-pricing sibling batching (W=8 target) on the design-D GEMV card [0xE4B711B0,0xE4B711D3]
  <repo>/reports/results/SIBLING_W_ON_GEMV_CARD_20260923.md
  Status: **INTERIM** — xsim gates, the fill-inclusive cycle census and the **W=4 L128 synth + place** are MEASURED;
- results:TAIL_FAST_20260923.md | 2026-09-23 | current | TAIL_FAST — the next TAIL image (0xE4B711CB candidate): TISEQ=1 + TREE_P=16 + SER_DEPTH=16384, full- [0xE4B71128,0xE4B711CA,0xE4B711CB]
  <repo>/reports/results/TAIL_FAST_20260923.md
  **Date** 2026-09-23 · **Fork** `verilog/forks/tail_fast_20260923` · **Submodule commits** `bdcf2bb` (source, the build export) +
- results:TREE_KPAR_20260923.md | 2026-09-23 | current | TREE_KPAR — where the TAIL's TREE time goes, and a bit-exact 35% cut (TISEQ=1 + TREE_P=16) [0xE4B711B0,0xE4B711CA]
  <repo>/reports/results/TREE_KPAR_20260923.md
  **Date** 2026-09-23 · **Fork** `verilog/forks/tree_kpar_20260923` (byte copy of `redesign_t0_tailnoattn_20260923`, the
- results:TWO_CARD_ARCHITECTURE_DECISION_20260923.md | 2026-09-23 | current | Two-card architecture decision for the next PDI pair — engine split vs tensor parallel vs layer spli [0xE4B711B0,0xE4B711C9,0xE4B711CA,0xE4B711D0...]
  <repo>/reports/results/TWO_CARD_ARCHITECTURE_DECISION_20260923.md
  **Date:** 2026-09-23 · **Scope:** READ-ONLY design research. No RTL edited, no build, no card access, nothing killed or
- results:UNI_GEMVL256_head_lab-host_PKG.md | 2026-09-23 | current | UNI_GEMVL256_head_lab-host_PKG.md
  <repo>/reports/results/UNI_GEMVL256_head_lab-host_PKG.md
  PACKAGE_OK WNS=10.140 URAM=1024 f158e5f16df396f0614b62e6b36f3de5  /mnt/data/eagle4_scratch/gemv_l256/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:UNI_REDESIGN_T0TNA_tail_lab-host_PKG.md | 2026-09-23 | current | UNI_REDESIGN_T0TNA_tail_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_T0TNA_tail_lab-host_PKG.md
  PACKAGE_OK WNS=3.374 URAM=640 cc0c8f78d715e48721a8b55964b245cc  /mnt/data/eagle4_scratch/redesign_t0tna/packaged_ip_oneflip_p2s_tail/eagle4_drafter_depthn_1_0/component.xml
- results:UNI_REDESIGN_T0W2_head_lab-host_PKG.md | 2026-09-23 | current | UNI_REDESIGN_T0W2_head_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_T0W2_head_lab-host_PKG.md
  PACKAGE_OK WNS=9.792 URAM=1072 4fb84b6f58d3c8b90301785de56712cf  /mnt/data/eagle4_scratch/redesign_t0w2/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:WHOLE_DRAFTER_FMAX_CENSUS_20260923.md | 2026-09-23 | current | Whole-drafter fmax census: what stops the drafter clock at 100 and 133 MHz (routed HEAD 0xE4B711C9,  [0xE4B711B0,0xE4B711C9,0xE4B711CA]
  <repo>/reports/results/WHOLE_DRAFTER_FMAX_CENSUS_20260923.md
  **Date:** 2026-09-23/24 · **Status:** IN PROGRESS — census jobs running on build-host-A (mesh class `timing_query_fullchip`).
- results:WINDOWS_WORKER_156_20260923.md | 2026-09-23 | current | Windows PC <ip> as an xsim / small-OOC worker — status 2026-09-23
  <repo>/reports/results/WINDOWS_WORKER_156_20260923.md
  - ping OK (~1 ms). TCP 22 OPEN; 3389, 5985 and 5986 closed.
- results:ACCEPT_COLLAPSE_ISOLATED_TO_HEADPAR8_20260922.md | 2026-09-22 | current | Accept 2.78 -> 1.00 ISOLATES TO ONE STRAP: `HEAD_PAR` 1 -> 8 [0xE4B71189,0xE4B711B0]
  <repo>/reports/results/ACCEPT_COLLAPSE_ISOLATED_TO_HEADPAR8_20260922.md
  **Method: on-disk differential of the two HEAD images' BUILT configs. No cards, no cold AC.**
- results:G8_ZERO_AR_REFUTED_20260922.md | 2026-09-22 | current | 🔺 RETRACTION — the "G8 zero-AR defect" is a MEASUREMENT ARTIFACT. G8 is not malfunctioning. [0xE4B71186]
  <repo>/reports/results/G8_ZERO_AR_REFUTED_20260922.md
  I reported twice, on 2026-09-22, that TAIL `0xE4B71186`'s G8 "launched, took all 512 candidates, saw
- results:HEADPAR8_SILICON_CONFIRMED_20260922.md | 2026-09-22 | current | 🟢 HEAD_PAR=8 CONFIRMED ON SILICON — the G5 slope divides by 7.95x, 0.6% from the cycle law [0xE4B71186,0xE4B71190]
  <repo>/reports/results/HEADPAR8_SILICON_CONFIRMED_20260922.md
  **Image:** HEAD `0xE4B71190` (R2 seam-export + HEAD_PAR=8, T0 lineage) @ `61:00.0`, paired with
- results:MD2_DEPTH_LOOP_DIAGNOSIS_20260922.md | 2026-09-22 | current | Why md=2 does not run on `0x1190`+`0x1186` — three mechanisms, none of them a broken image
  <repo>/reports/results/MD2_DEPTH_LOOP_DIAGNOSIS_20260922.md
  All four gates that fired tonight fired **correctly**. The pair is **healthy**: after every experiment
- results:MD2_UNBLOCKED_20260922.md | 2026-09-22 | current | 🟢 md=2 RUNS ON THE PAIR — the ABI-latch fix works, and the TAIL was never depth-limited [0xE4B71186,0xE4B71190]
  <repo>/reports/results/MD2_UNBLOCKED_20260922.md
  Cold AC at 03:12. **Armed BEFORE the first `AP_START`, exactly as `PERSIST_ABI_LATCH_20260922.md`
- results:PERSIST_ABI_LATCH_20260922.md | 2026-09-22 | current | 🔴 ON A PERSIST IMAGE THE md=1/md=2 OUT-DMA ABI IS FIXED AT THE FIRST `AP_START` AFTER BOOT
  <repo>/reports/results/PERSIST_ABI_LATCH_20260922.md
  **And I spent this boot's one chance on md=1.** Ring cleared, phase2 retried as asked — it fails for a
- results:PER_PHASE_TAIL_ANSWER_20260922.md | 2026-09-22 | current | 🟢 THE PER-PHASE DUMP — G8 IS 61.6% OF THE TAIL, AT 260x ITS OWN ARITHMETIC
  <repo>/reports/results/PER_PHASE_TAIL_ANSWER_20260922.md
  One-shot post-cold-AC probe, **exactly one item**, md=1, ctx=1, K_C=256, both cards.
- results:SESSION_CONTEXT_BACKUP_20260922.md | 2026-09-22 | current | SESSION CONTEXT BACKUP 2026-09-22 — post-cold-AC bring-up, HEAD_PAR=8 confirmed, md=2 foreclosed [0xE4B71186,0xE4B71190]
  <repo>/reports/results/SESSION_CONTEXT_BACKUP_20260922.md
  Session ran 2026-09-21 22:5x → 2026-09-22 03:0x on **lab-host only** (user: "do not use build-host-M").
- results:SGLANG_PAIR_1190_ACCEPT_COLLAPSE_20260922.md | 2026-09-22 | current | SGLang RUNS on the 0x1190 pair — first time ever — but ACCEPT COLLAPSES TO 1.00
  <repo>/reports/results/SGLANG_PAIR_1190_ACCEPT_COLLAPSE_20260922.md
  **`EXIT=0`.** The pair served 5 humaneval requests end-to-end under SGLang. **This is the first time
- results:SGLANG_PAIR_1190_ATTEMPT_20260922.md | 2026-09-22 | current | SGLang serving on HEAD `0x1190` + TAIL `0x1186` — path reaches silicon, blocked on a SoA shape misma [0xE4B71185,0xE4B71186,0xE4B71190]
  <repo>/reports/results/SGLANG_PAIR_1190_ATTEMPT_20260922.md
  **Three runs, three distinct failures, all diagnosed. Neither card wedged; both healthy afterwards.**
- results:T0_CNTCLR_GATE_VERDICT_20260922.md | 2026-09-22 | current | T0 `cnt_clr` gate — RUN FOR THE FIRST TIME, 2026-09-22 (GREEN, two-sided) [0xE4B711B0]
  <repo>/reports/results/T0_CNTCLR_GATE_VERDICT_20260922.md
  **Status: CERTIFIED PRE-SILICON.** Until this run the instrument that all three VM TAIL builds
- results:T0_GOLDEN_INHERITANCE_20260922.md | 2026-09-22 | current | Do the T0 `cnt_clr` arms still match the GPU goldens? — datapath identity argument [0xE4B71189,0xE4B711B0]
  <repo>/reports/results/T0_GOLDEN_INHERITANCE_20260922.md
  **USER REQUIREMENT (2026-09-22): "by working with sglang it also needs to match the gpu goldens
- results:UNI_REDESIGN_T0_head_lab-host_PKG.md | 2026-09-22 | current | UNI_REDESIGN_T0_head_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_T0_head_lab-host_PKG.md
  PACKAGE_OK WNS=9.792 URAM=1072 4e6313c4d6aee0407e9c246be8fb152c  /mnt/data/eagle4_scratch/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:VM_TAIL_CAMPAIGN_20260922.md | 2026-09-22 | current | OVERNIGHT VM CAMPAIGN — three TAIL PDIs on the Oleg build VM [0xE4B71111,0xE4B71118,0xE4B71185,0xE4B71186...]
  <repo>/reports/results/VM_TAIL_CAMPAIGN_20260922.md
  Launched 2026-09-22 ~04:2x on user instruction ("drive autonomously towards multiple pdi images…
- results:WHY_THE_PAIR_IS_SLOWER_20260922.md | 2026-09-22 | current | Why two FPGAs are SLOWER than one — regime first, then structure
  <repo>/reports/results/WHY_THE_PAIR_IS_SLOWER_20260922.md
  **Question (user 2026-09-22): "why two fpga made it slower? we have 120 tps with a single fpga."**
- results:ALT_BUILD_MACHINE_OLEG_VM_20260921.md | 2026-09-21 | current | Alternative build machine (Oleg VM / build-host-M) + Tyler runnable reference — 2026-09-21 [0xE4B71189,0xE4B7118A,0xE4B7118B,0xE4B7118C...]
  <repo>/reports/results/ALT_BUILD_MACHINE_OLEG_VM_20260921.md
  Owner: lab-host Claude Code, FORK of the drafter orchestrator (user, 2026-09-21 ~01:25 EDT: "fork yourself to do
- results:ARCH_FOR_CHIP_DESIGN_20260921.md | 2026-09-21 | current | EAGLE4 drafter — microarchitecture reference for chip design [0xE4B71185]
  <repo>/reports/results/ARCH_FOR_CHIP_DESIGN_20260921.md
  **Source of truth:** fork `pipe_oneflip_g7sp` at
- results:GPU_PHASE_TENSOR_RTL_GATES_20260921.md | 2026-09-21 | current | GPU-captured per-phase / per-tensor RTL validation — R3 D.2 fork, 2026-09-21 [0xE4B71193]
  <repo>/reports/results/GPU_PHASE_TENSOR_RTL_GATES_20260921.md
  **Scope.** Run the *existing* GPU-capture parity suite against the **newest RTL** — the R3 KV-only
- results:HOST_W1W2_HARDENING_20260921.md | 2026-09-21 | current | HOST-SIDE W1/W2 HARDENING — implementation of the R3 D.2 audit's host change list [0xE4B71185,0xE4B71189,0xE4B7118A,0xE4B7118B...]
  <repo>/reports/results/HOST_W1W2_HARDENING_20260921.md
  **Date** 2026-09-21 · **Scope** host Python/shell only, MOCK-ONLY · **Input** `reports/results/R3_D2_COORDINATION_AUDIT_20260921.md` (H1, H3c, H3e + both OVERALL lists)
- results:PAIR_ACCURACY_AND_ATTRIBUTION_20260921.md | 2026-09-21 | current | Pair accuracy + step attribution under SGLang — 2026-09-21 (fork: silicon + GPU track) [0xE4B71185,0xE4B71186,0xE4B71192,0xE4B71193]
  <repo>/reports/results/PAIR_ACCURACY_AND_ATTRIBUTION_20260921.md
  Pair RUNNING at start: HEAD `0xE4B71185`@61:00.0 / TAIL `0xE4B71186`@41:00.0 (reg 0x100 read by the sampler).
- results:PAIR_MD1_RELAY_POSTAC_20260921.md | 2026-09-21 | current | md=1 TWO-CARD RELAY PASSES ON SILICON — first run after the 2026-09-21 cold AC [0xE4B71186,0xE4B71190]
  <repo>/reports/results/PAIR_MD1_RELAY_POSTAC_20260921.md
  **Pair:** HEAD `0xE4B71190` (R2 seam-export + **HEAD_PAR=8**, T0 lineage) @ `61:00.0` ·
- results:PAIR_STATE_PROFILE_20260921.md | 2026-09-21 | current | Depth-N pair — per-phase time on LIVE silicon by FSM-state sampling (2026-09-21) [0xE4B71185,0xE4B71186]
  <repo>/reports/results/PAIR_STATE_PROFILE_20260921.md
  Images RUNNING (reg 0x100, read by the sampler itself): HEAD `0xE4B71185`@61:00.0, TAIL `0xE4B71186`@41:00.0.
- results:POSTAC_BRINGUP_20260921.md | 2026-09-21 | current | POST-COLD-AC BRING-UP 2026-09-21 22:5x — BOTH CARDS UP, BOTH BUILD-IDs CONFIRMED [0xE4B71138,0xE4B71139,0xE4B71186,0xE4B71190]
  <repo>/reports/results/POSTAC_BRINGUP_20260921.md
  User performed the physical cold AC after the UPDATE 618 flash pair. Box `uptime` 17 min at first
- results:R1_HP4_ACCURACY_GATE_20260921.md | 2026-09-21 | current | R1 HEAD_PAR=4 ACCURACY GATE — full-drafter HEAD xsim arm vs the HEAD_PAR=1 reference [0xE4B7118D,0xE4B7118E,0xE4B71191]
  <repo>/reports/results/R1_HP4_ACCURACY_GATE_20260921.md
  **Date:** 2026-09-21 · **Host:** Oleg build VM `<build-vm>`
- results:R1_HP8_ACCURACY_GATE_20260921.md | 2026-09-21 | current | R1 ACCURACY GATE — `HEAD_PAR=8` (G5_HPS=4) vs the silicon-proven `HEAD_PAR=1`, xsim, at the `0xE4B71 [0xE4B7118C]
  <repo>/reports/results/R1_HP8_ACCURACY_GATE_20260921.md
  **2026-09-21 · xsim only.** Sandbox `/mnt/data/eagle4_scratch/r1hp8_gate/` (create-only). No card, no GPU, no build, no
- results:R3_D2_COORDINATION_AUDIT_20260921.md | 2026-09-21 | current | R3 (W1 re-arm + soft abort, W2 PERSIST KV-only item) — READ-ONLY COORDINATION AUDIT [0xE4B71192,0xE4B71193]
  <repo>/reports/results/R3_D2_COORDINATION_AUDIT_20260921.md
  **Date** 2026-09-21 · **Auditor role** read-only pair-auditor (did not write the change) · **Method** reading only — no Vivado/xsim/xelab/xvlog run (6 GB RAM cap); prior gate LOGS were read, never re-
- results:R4_G8NCH4_COORDINATION_AUDIT_20260921.md | 2026-09-21 | current | R4 G8_NCH=4 — READ-ONLY coordination audit (2026-09-21) [0xE4B71186]
  <repo>/reports/results/R4_G8NCH4_COORDINATION_AUDIT_20260921.md
  **Audited:** `verilog/forks/redesign_r4_g8nch4_20260921/` = submodule commit `05080ed` (parent `ca39b44`); parent gitlink
- results:R4_G8NCH4_PHASE1_20260921.md | 2026-09-21 | current | R4 phase 1 — G8 exact-rescore gather, 2 → 4 HBM read channels: fork + RTL + xsim proof (2026-09-21) [0xE4B71186,0xE4B71192]
  <repo>/reports/results/R4_G8NCH4_PHASE1_20260921.md
  **Scope actually exercised:** xvlog / xelab / xsim only. No Vivado synthesis, no packaging run, no BD edit on any AVED clone, no
- results:TAIL_R4_LEVER_INVENTORY_20260921.md | 2026-09-21 | current | TAIL R4 — accuracy-neutral speed-lever inventory for the depth-N pair's TAIL (2026-09-21) [0xE4B7115B,0xE4B71167,0xE4B71186,0xE4B71192]
  <repo>/reports/results/TAIL_R4_LEVER_INVENTORY_20260921.md
  **READ-ONLY analysis.** No Vivado / xsim / xelab run, no RTL or script edited, nothing killed, no card touched, no commit.
- results:UNI_REDESIGN_R1HP8_head_lab-host_PKG.md | 2026-09-21 | current | UNI_REDESIGN_R1HP8_head_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_R1HP8_head_lab-host_PKG.md
  PACKAGE_OK WNS=10.051 URAM=1072 dd5ad69d4ce131ff47fef3ed272b6b67  /mnt/data/eagle4_scratch/redesign_r1hp8/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:UNI_REDESIGN_R2HP8_head_lab-host_PKG.md | 2026-09-21 | current | UNI_REDESIGN_R2HP8_head_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_R2HP8_head_lab-host_PKG.md
  PACKAGE_OK WNS=10.069 URAM=1072 78def5d4d7c108dbe46764ec2d731716  /mnt/data/eagle4_scratch/redesign_r2hp8/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:UNI_REDESIGN_R3HP8_head_lab-host_PKG.md | 2026-09-21 | current | UNI_REDESIGN_R3HP8_head_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_R3HP8_head_lab-host_PKG.md
  PACKAGE_OK WNS=10.069 URAM=1072 aae9c29a04d8f047d3f17973afac7d12  /mnt/data/eagle4_scratch/redesign_r3/packaged_ip_oneflip_p2s_head/eagle4_drafter_depthn_1_0/component.xml
- results:UNI_REDESIGN_R3TAIL_tail_lab-host_PKG.md | 2026-09-21 | current | UNI_REDESIGN_R3TAIL_tail_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_R3TAIL_tail_lab-host_PKG.md
  PACKAGE_OK WNS=3.374 URAM=816 54e59426cdac11c7b29309f84464b011  /mnt/data/eagle4_scratch/redesign_r3/packaged_ip_oneflip_p2s_tail/eagle4_drafter_depthn_1_0/component.xml
- results:UNI_REDESIGN_T0_tail_lab-host_PKG.md | 2026-09-21 | current | UNI_REDESIGN_T0_tail_lab-host_PKG.md
  <repo>/reports/results/UNI_REDESIGN_T0_tail_lab-host_PKG.md
  PACKAGE_OK WNS=3.374 URAM=816 a13923296eddd9e41ba4e7fd55b62a5c  /mnt/data/eagle4_scratch/redesign_t0/packaged_ip_oneflip_p2s_tail/eagle4_drafter_depthn_1_0/component.xml
- results:VM_BUILD_ARCHITECTURE_20260921.md | 2026-09-21 | current | Architecture record — the redesign HEAD images built on the Oleg VM (and their lab-host/build-host-M sibli [0xE4B71185,0xE4B7118D,0xE4B7118E,0xE4B7118F...]
  <repo>/reports/results/VM_BUILD_ARCHITECTURE_20260921.md
  **Dated 2026-09-21 14:00 EDT. For Carrick's review.** Part 1 (this file) = what each image IS: role, straps, lineage, resources,
- results:VM_BUILD_ARCHITECTURE_PART2_ENGINE_AXI_20260921.md | 2026-09-21 | current | Part 2 — engine- and module-level mechanism + AXI/NoC/register wiring of the redesign HEAD images
  <repo>/reports/results/VM_BUILD_ARCHITECTURE_PART2_ENGINE_AXI_20260921.md
  **Dated 2026-09-21. For Carrick's review. Companion to `VM_BUILD_ARCHITECTURE_20260921.md` (Part 1 = what the
- results:KC_L0_DEPTH2_SILICON_20260919.md | 2026-09-19 | current | K_C x depth-2 golden identity on the LIVE 0x1185/0x1186 pair — 2026-09-19 [0xE4B71185,0xE4B71186]
  <repo>/reports/results/KC_L0_DEPTH2_SILICON_20260919.md
  **Verdict: the depth-N pair is CORRECT at depth 2. The prior "L0 divergence" was a comparator
- results:PAIR_TIME_BUDGET_20260919.md | 2026-09-19 | current | Depth-N pair time budget — measured under real SGLang, 2026-09-19 [0xE4B71185,0xE4B71186]
  <repo>/reports/results/PAIR_TIME_BUDGET_20260919.md
  Images RUNNING (reg 0x100): HEAD `0xE4B71185`@`0000:61:00.0`, TAIL `0xE4B71186`@`0000:41:00.0`.
- results:QSFP_1185_DEPTH2_SGLANG_TESTING_20260919.md | 2026-09-19 | current | HEAD1185 / TAIL1186: depth-2 verification and integrated SGLang [0xE4B71185,0xE4B71186]
  <repo>/reports/results/QSFP_1185_DEPTH2_SGLANG_TESTING_20260919.md
  Measured on lab-host, 2026-09-19. Resident images remained 0xE4B71185 / 0xE4B71186. Direct user authority: further silicon testing, including integrated SGLang TPS. No image load, reset, queue reconfig
- results:QSFP_1185_GPU_OP_RTL_VERIFICATION_20260919.md | 2026-09-19 | current | HEAD1185 operator RTL verification against OSPREY captures
  <repo>/reports/results/QSFP_1185_GPU_OP_RTL_VERIFICATION_20260919.md
  Date: September 19, 2026. Independent Codex QSFP verification in response to the user's direct request. No hardware accesses, flash, resets, builds, ticket ownership changes, or changes to the drafter
- results:QSFP_1185_SILICON_OP_DEBUG_20260919.md | 2026-09-19 | current | HEAD1185 / TAIL1186 silicon tensor localization — measured result [0xE4B71185,0xE4B71186]
  <repo>/reports/results/QSFP_1185_SILICON_OP_DEBUG_20260919.md
  Updated 2026-09-19 02:03 EDT. Authenticated access restored; first observed failing operator localized to G4 Q and corrected by restoring valid packed scales. Root HEAD tensors and root top-8 now pass
- results:QSFP_1185_SPEED_AUDIT_20260919.md | 2026-09-19 | current | Why the current diagnostic path is 11.58 TPS versus the historical 132.13 TPS [0xE4B71167]
  <repo>/reports/results/QSFP_1185_SPEED_AUDIT_20260919.md
  Offline analysis of existing measured logs, 2026-09-19. No hardware access, build changes, flash, reset, power action, or external publication. The historical 132.13 TPS result is present in its origi
- results:QSFP_CODEX_DEPTHN_DELIVERY_AUDIT_20260919.md | 2026-09-19 | current | Codex depth-N delivery audit and resumed host debugging [0xE4B71185,0xE4B71186]
  <repo>/reports/results/QSFP_CODEX_DEPTHN_DELIVERY_AUDIT_20260919.md
  Recorded 2026-09-19T00:55:16-04:00. This is an independent offline audit under the user's direct request; the drafter's ongoing silicon investigation retains ownership.
- results:ANOMALY_DEPTHN_RELAY_20260918.md | 2026-09-18 | current | ANOMALY — Depth-N candidate-feedback RELAY datapath (HEAD 0xE4B71185 / TAIL 0xE4B71186) [0xE4B71185,0xE4B71186]
  <repo>/reports/results/ANOMALY_DEPTHN_RELAY_20260918.md
  **Verdict (one line):** CLEAN — the region-3 `for_reasoning` relay shows **NO new latent silicon-fatal
- results:DEPTHN_BRINGUP_DRESS_REHEARSAL_20260918.md | 2026-09-18 | current | DEPTH-N PAIR (HEAD 0x1185 / TAIL 0x1186) — POST-COLD-AC BRING-UP DRESS REHEARSAL [0xE4B71185,0xE4B71186]
  <repo>/reports/results/DEPTHN_BRINGUP_DRESS_REHEARSAL_20260918.md
  **Date:** 2026-09-18 · **Author:** [redacted-addr] (host-only, read-only/create-only rehearsal agent)
- results:DEPTHN_DEPTH3_GOLDEN_20260918.md | 2026-09-18 | current | Depth-3 OSPREY MODEL-TRUTH golden (Track 1 depth-N prep) — 2026-09-18
  <repo>/reports/results/DEPTHN_DEPTH3_GOLDEN_20260918.md
  **READY: depth-3 OSPREY golden captured, md5 `cddfb84050faa8ca0b49d788b1e52b87`, depth-1 control 8/8, depth-2 sub-tree matches `ddc47efaab2393a5a8a9b38f48da84e0`.**
- results:DEPTHN_DEPTH_SWEEP_VALIDATE_20260918.md | 2026-09-18 | current | Depth-2..15 one-command SWEEP validator — Track 1 depth-N (2026-09-18)
  <repo>/reports/results/DEPTHN_DEPTH_SWEEP_VALIDATE_20260918.md
  **READY: sweep validator authored, selftest PASS (15 goldens md2..md15 md5-verified, comparison GREEN
- results:DEPTHN_GOLDEN_LADDER_20260918.md | 2026-09-18 | current | Depth-4..8 OSPREY MODEL-TRUTH golden ladder (Track 1 depth-N prep) — 2026-09-18
  <repo>/reports/results/DEPTHN_GOLDEN_LADDER_20260918.md
  **READY: depth-4..8 OSPREY goldens captured (md5 md4 `ff03e48ef230b12e349e3fd543b410af`,
- results:FSMSPLIT_BATCHID_0x116B_BUILD_20260918.md | 2026-09-18 | current | FSMSPLIT_BATCHID 0xE4B7116B single-card PDI build — 2026-09-18 [0xE4B7116B]
  <repo>/reports/results/FSMSPLIT_BATCHID_0x116B_BUILD_20260918.md
  Single-card PDI = the 0x1167 record image + the ONE new strap **TREE_BATCH_ID=1**
- results:FSMSPLIT_BATCH_IDENTITY_AUDIT_20260918.md | 2026-09-18 | current | FSM_SPLIT batch-identity — coordination audit (Track 3, RCA §5c) [0xE4B71118]
  <repo>/reports/results/FSMSPLIT_BATCH_IDENTITY_AUDIT_20260918.md
  **Date** 2026-09-18 · **Type** READ-ONLY PAIR AUDIT (gates the FS=1+TREE_BATCH_ID=1 single-card build) ·
- results:FSMSPLIT_BATCH_IDENTITY_DESIGN_20260918.md | 2026-09-18 | current | FSM_SPLIT batch-identity — design study + prototype (Track 3, RCA §5c)
  <repo>/reports/results/FSMSPLIT_BATCH_IDENTITY_DESIGN_20260918.md
  **Date** 2026-09-18 · **Type** DESIGN STUDY / CANDIDATE (no build, no flash, no card) ·
- results:FSMSPLIT_P2_REEXAM_20260918.md | 2026-09-18 | current | Track-3 (FSM_SPLIT) re-exam: reconciling "D594-P2 UNMET" vs the 0x1155/0x1167 accept-collapse — GO/N [0xE4B71155]
  <repo>/reports/results/FSMSPLIT_P2_REEXAM_20260918.md
  **Date** 2026-09-18 · **Scope** Reconcile the two open FSM_SPLIT findings and deliver a firm verdict for the
- results:HEAD_KPAR32_OOC_20260918.md | 2026-09-18 | current | HEAD K_PAR 16→32 — OOC SYNTHESIS OF THE SCALED GEMV LANE ARRAY (task #66) [0xE4B71185]
  <repo>/reports/results/HEAD_KPAR32_OOC_20260918.md
  **Date:** 2026-09-18 · **Scope:** READ-ONLY / CREATE-ONLY. New diagnostic fork copied (not moved);
- results:HEAD_KPAR32_SCALEDLANE_20260918.md | 2026-09-18 | current | HEAD K_PAR 16→32 — K_PAR-PARAMETRIC SCALED GEMV LANE (task #66) [0xE4B71185]
  <repo>/reports/results/HEAD_KPAR32_SCALEDLANE_20260918.md
  **Date:** 2026-09-18 · **Scope:** CREATE-ONLY. New fork copied (not moved); new lane + broken NEG-control
- results:HEAD_KPAR32_SCALEDLANE_AUDIT_20260918.md | 2026-09-18 | current | HEAD K_PAR 16→32 SCALED-LANE — COORDINATION AUDIT (bit-exact methodology)
  <repo>/reports/results/HEAD_KPAR32_SCALEDLANE_AUDIT_20260918.md
  **Role:** READ-ONLY pair-auditor for the K_PAR-parametric SCALED body-GEMV lane (task #66).
- results:PIPELINE_0x1169_PREFLASH_READINESS_20260918.md | 2026-09-18 | current | Pipeline 0x1169 (two-candidate) — Pre-Flash Readiness (2026-09-18) [0xE4B71169]
  <repo>/reports/results/PIPELINE_0x1169_PREFLASH_READINESS_20260918.md
  Read-only agent, host lab-host. Scratch: `/mnt/data/preflash_1169_scratch/`.
- results:PIPELINE_0x1169_TIMING_RCA_20260918.md | 2026-09-18 | current | Pipeline 0x1169 — Post-Route Timing RCA + Retry Verdict (2026-09-18)
  <repo>/reports/results/PIPELINE_0x1169_TIMING_RCA_20260918.md
  **Author:** host-side RCA agent (READ-ONLY / CREATE-ONLY). No build/flash/kill/rm performed.
- results:SESSION_STATE_FULL_20260918.md | 2026-09-18 | current | EAGLE4 SESSION STATE — FULL, WITH PATHS (2026-09-18 23:06) [0xE4B71167,0xE4B71169,0xE4B7116B,0xE4B71185...]
  <repo>/reports/results/SESSION_STATE_FULL_20260918.md
  Single-page index of everything live/produced this session, every path absolute or repo-relative.
- results:BODYWNS_0x1185_RESOLUTION_20260917.md | 2026-09-17 | current | BODYWNS 0xE4B71185 (HEAD depth-N) Quarantine Resolution — 2026-09-17 [0xE4B71185,0xE4B71186]
  <repo>/reports/results/BODYWNS_0x1185_RESOLUTION_20260917.md
  **Author:** [redacted-addr] · read-only/create-only host agent
- results:CODEX_SINGLE_D1_COLLAPSE_ANSWER_DRAFT_20260917.md | 2026-09-17 | current | DRAFT ANSWER — Codex "single-card depth-1 acceptance collapse" (0xe4b71167) [0xE4B71128,0xE4B71167]
  <repo>/reports/results/CODEX_SINGLE_D1_COLLAPSE_ANSWER_DRAFT_20260917.md
  **Status: DRAFT for orchestrator review. NOT posted to the relay.**
- results:DEPTHN_BUILD_READINESS_20260917.md | 2026-09-17 | current | Depth-N candidate-feedback pair — BUILD-READINESS (2026-09-17) [0xE4B7113B,0xE4B7113C]
  <repo>/reports/results/DEPTHN_BUILD_READINESS_20260917.md
  **Mission (user Track 1).** Take the depth-N candidate-feedback DESIGN
- results:DEPTHN_CANDIDATE_FEEDBACK_DESIGN_20260917.md | 2026-09-17 | current | Depth-N candidate feedback for the one-flip HEAD/TAIL pair — RTL + host design (2026-09-17)
  <repo>/reports/results/DEPTHN_CANDIDATE_FEEDBACK_DESIGN_20260917.md
  **Mission (user Track 1).** Give the one-flip HEAD/TAIL pair a multi-level tree by transmitting the
- results:DEPTHN_DEPTH2_GOLDEN_20260917.md | 2026-09-17 | current | Depth-2 OSPREY golden frontier for the depth-N pair (0x1185/0x1186)
  <repo>/reports/results/DEPTHN_DEPTH2_GOLDEN_20260917.md
  **Date:** 2026-09-17 · **Author:** [redacted-addr] (GPU-side agent; create-only, no card/flash/build/kill/rm)
- results:DEPTHN_PAIR_ABI_CONTRACT_20260917.md | 2026-09-17 | current | Depth-N Paired-Card Host Register / ABI Contract (RTL-derived, authoritative)
  <repo>/reports/results/DEPTHN_PAIR_ABI_CONTRACT_20260917.md
  **Date:** 2026-09-17
- results:DEPTHN_PAIR_BUILD_LAUNCH_20260917.md | 2026-09-17 | current | Depth-N md=2 PAIR — BUILD LAUNCH (2026-09-17) [0xE4B71185,0xE4B71186]
  <repo>/reports/results/DEPTHN_PAIR_BUILD_LAUNCH_20260917.md
  **Verdict: LAUNCHED (report-only gated, flashable-but-QUARANTINED; NO flash — flash is USER).**
- results:DEPTHN_PAIR_FLASH_BRINGUP_RUNBOOK_20260917.md | 2026-09-17 | current | DEPTH-N PAIR (HEAD 0x1185 / TAIL 0x1186) — FLASH + BRING-UP RUNBOOK [0xE4B71138,0xE4B71139,0xE4B7113B,0xE4B7113C...]
  <repo>/reports/results/DEPTHN_PAIR_FLASH_BRINGUP_RUNBOOK_20260917.md
  **Date:** 2026-09-17 · **Author:** [redacted-addr] (read-only/create-only host agent)
- results:DEPTHN_PAIR_RUNTIME_DRIVER_20260917.md | 2026-09-17 | current | Depth-N Paired-Card Runtime Driver — create-only deliverable (HEAD 0x1185 / TAIL 0x1186) [0xE4B71185,0xE4B71186]
  <repo>/reports/results/DEPTHN_PAIR_RUNTIME_DRIVER_20260917.md
  **Date:** 2026-09-17 · **Author:** [redacted-addr] (create-only host agent; no card, no flash, no build, no kill, no rm)
- results:DEPTHN_RELAY_RTL_VERIFY_20260917.md | 2026-09-17 | current | Depth-N candidate-feedback (TAIL→HEAD) RELAY — RTL-present verification for HEAD `0xE4B71185` [0xE4B71185]
  <repo>/reports/results/DEPTHN_RELAY_RTL_VERIFY_20260917.md
  **Date** 2026-09-17 · **Author** read-only host-side verification agent (lab-host; no card, no build, no kill, no rm)
- results:FPGA_PIPELINE_TWO_CANDIDATE_DESIGN_20260917.md | 2026-09-17 | current | TWO-CANDIDATE FPGA PIPELINE — a card-pair design that raises single-stream tok/s by drafting two ind [0xE4B71167]
  <repo>/reports/results/FPGA_PIPELINE_TWO_CANDIDATE_DESIGN_20260917.md
  **Scope.** READ-ONLY design. No card access, no build, no flash, no reset, no process killed, nothing removed. Create-only.
- results:FSMSPLIT_FIX_STATE_20260917.md | 2026-09-17 | current | FSM_SPLIT DUPLICATE-FRONTIER — CONSOLIDATED FIX STATE (2026-09-17) [0xE4B71167]
  <repo>/reports/results/FSMSPLIT_FIX_STATE_20260917.md
  **Track-3 companion (host + xsim only). REPORT-ONLY. No card, no flash, no reset, no kill, no `rm`.**
