# Will the demo images work on silicon? Functional-readiness audit, 2026-11-10 (07:15–09:00Z)

Read-only audit by a box1-two-v2 worker: no card, no card lock, no build launched, nothing killed or removed. talon2 was reached
for file listings and file staging only. Program 1 = the ONE Llama-3.2-3B drafter (SET 1 on 7 cards, SET 2 on 4 cards); Qwen is noted
where it differs. Status words: **PROVEN** (an artifact shows it), **PARTIAL** (shown at a different layer/shape than what ships),
**MISSING** (nothing shows it). [M] measured here, [R] read from RTL/kit, [C] reasoning.

## Answer first
- **SET 1 (6-card TP4: ATT B4 on 17/3d/bd/e1 + MLPV 0x134D on 50/63) was not runnable at 07:15Z for host reasons, not RTL reasons.**
  There was no boot root for any of the six cards, no 6-card loop/golden/serve path, and the kit's identity check refused B4 outright.
  Closed during this audit: the MLPV boot root (built, proven, staged on both hosts), the identity collision (one line), a 6-card mock
  that runs main's new Session end to end, the 4 KB boot-burst scan. Still open: the 6-card serve path, a comparator for correctness
  (the 4-card golden is the wrong one), 0x134D's PDI adjudication, and three RTL-level "never simulated together" rows.
- **SET 2 (4-card BATCH_MODE): no flashable pair exists.** MLPV 0x134C is routed (+0.014 / WHS +0.008 with margin, PDI not written,
  not adjudicated). Its intended ATT partner C4 0xE4B713C4 is **not launched**. The only silicon-proven multi-stream results are on
  other images (0x1334 static 2 banks: 647 aggregate; F3t + batched verify + HBM replay: 730). Nothing in the host kit writes
  0x0FC / 0x240. 0x134C in mode 0 is a drop-in 0x1342 (xsim identity), so its realistic value tonight is "0x1342 that can later take mode 2".
- **New silicon bug class (M9 4 KB burst, item a0): no boot root of either set crosses a page.** The scanner reproduces the real
  failure on the Q5120 v1 kit at the silicon address.

## a0. M9 boot-DMA bursts crossing 4 KB [M]
Loader: `weight_dma_loader.sv` md5 `e93db7be` is the SAME file in the 0x134D fork (`verilog/forks/fourcard_4c1_mlpv_tp4_20261109/src_full/top/`)
as in the Q5120 fork that wedged card 63: fixed INCR bursts of `BOOT_BURST_LEN` 64 beats x 4 B from the CSR base, no 4 KB split
(`:67-70`, `this_burst_beats`, `S_ARREQ`; `eagle4_config.svh:468`). Tool: `scripts/4c1/m9_4kb_scan.py` (new; replays every `*.ops`,
models `f_total_words` and the burst walk; `--selftest` PASS, GREEN 3 / RED 4 / WARN 1).

| boot root | rank dirs | verdict | first offending address |
|---|---|---|---|
| talon2 `q5120_20261110/boot/q5120_boot_v1/0xE4B71349` rank0, rank1 (the kit that wedged 63) | 2 | **FAIL** (the RED on real data) | 120 of 518 loads; `boot.ops:3193` burst `0x42011e0f30` +256 B = the silicon address |
| talon2 `q5120_boot_v2` | all | PASS | — |
| box1 `~/eagle4_artifacts/4c1_boot` + `4c1_boot_qwen27b` (+ v2/v3 roots): 0x1342, 0x13F9, 0x1334, 0x1339, 0x133C, … | 287 | PASS | — |
| talon2 `qwen/artifacts/4c1_boot_llama`, `4c1_boot_qwen27b`, `~/eagle4_artifacts/4c1_boot` | 173 | PASS | — |
| **0x134D TP4 MLPV** `~/eagle4_artifacts/4c1_boot_tp4/0xE4B7134D` (box1 + talon2) | 2 | **PASS** (2 M9 loads each, bases 4 KB aligned) | RED: same `boot.ops` with the G9 scale base at `…f30` → FAIL `0x4200000f30` |
| **B4 / B1 / B5 ATT** `…/4c1_boot_tp4/0xE4B713B{4,1,5}/rank0..3` (junior's generator) | 12 | **PASS** (12 M9 loads each = 6 scale + 6 D9ZERO) | — |
| 0x134C / 0x1356 batch MLPV, C4 / BA / BB | 0 | **no boot root exists** (MISSING, not PASS) | — |

Why the Llama roots are safe by construction [R]: `gen_4c1_boot_streams.py:414` and `c4_tp4_weights.py:375` issue ONE M9 load per slice
from a base advanced by `ceil(nbytes / 4096) * 4096`. The risk appears only with per-lane loads (the Q5120 G12 case). RTL follow-up
(owner: whoever cuts the next MLPV/ATT): split the burst at 4 KB or refuse an unaligned base.

## SET 1 — 6-card TP4 (ATT B4 0xE4B713B4 / alternates B1, B5; MLPV 0xE4B7134D)

| # | item | status | evidence | fix / owner | hours |
|---|---|---|---|---|---|
| a1 | ATT→MLPV frame contract of B4 == the AD2 contract 0x134D was built against | **PROVEN** [M] | Stage `common/` of B4 == B1 (AD2 stack) file for file except one comment line in `pa_overlay_d2.tcl`: `loop_seq_d2.sv` `076a764b` (= the contract's md5), `link_dma_d2_bd.v` `746db910`, `link_dma_eng_lp_v2.sv` `4ed646cf`, `tp_link4_tp4_bd.v` `6d63efcf`, `tp_ar4_hc.sv` `73902acc`, `att_g6_tp.sv` `871693e1` (`/mnt/data/eagle4_scratch/tp4_stack_20261109/stages/lutg_d9_tp4stk_{B1,B4,B5}_stage`). B4 `GATE_VERDICTS` `:91` BELL THREEC_A=2 22,528 B, `:27` SG_ATT_TP_RANKS 4, `:73` ATTN_N_HEADS 8, `:355` ATT4_BD_VALIDATED tp4hc=1 | none | 0 |
| a2 | B4-only deltas vs AD2 | **PROVEN, host consequence** | LANES 512 (AD2 256), RET_SER 1 R8, SG_ATT_TP4_G4N 3072 (AD2 2304), G6_REPLAY_DB 0. All array-internal. r5 quarter 128 x 32 B @+0, h1 @+0x4000, `a` @+0x11800, stride 0x14000 unchanged. **B4 needs lanes-512 rank dirs, B1/B5 lanes-256: not interchangeable** (junior generated both) | A/B swap C3 = other boot root | 0 |
| a3 | MLPV side of the contract (ring A/B, quarter split, join, fan-out) | **PROVEN in parts** | `SEVENCARD_MLPV_0x134D_20261109.md` §5b link gate GREEN 6/6, REDs 4/4; §E2E | see b | — |
| b1 | 0x134D e2e TP4 arm `g2tp4_md3_bl10` recorded **NOT_AS_REQUIRED** | **PROVEN GREEN after re-adjudication** [M] | The 2 "mismatches" are `tfd/stages.txt` on r0 and r1, and only the `ING0` digest lines differ (1,024 per rank). They are exactly the expected permutation (local cells 0..127 = quarter m, 128..255 = quarter m+2; 8/8 items, both ranks; 0 poison cells ingested). Every other line (G9LO, R7, G11, GLM, CAND, TREE, RESC), all `a_k*.bin`, both SoA, feed, g12, level window and `cycles.txt` are byte-identical to the 0x1342 parent run. REDs r1wr / r2dup / r3teeth mismatch on 28 files | comparator should exclude ING0 on a strap-1 arm (box1-two-v2) | 0.2 |
| b2 | Full chain: 4 ATT ranks (B4 RTL) + real hypercube + real link + 2 MLPV (0x134D RTL) in ONE simulation | **MISSING** | Three disjoint islands: (i) ATT 4-rank world on AC = B4 RTL (124/125 files), ideal 4-way SUM, no MLPV, 4 items; (ii) link gate `tb_solo`: 2 wave-D HEADs + the merged TAIL `b086b6b0` (not AD2's `076a764b` HEAD); (iii) MLPV pair world fed by prebuilt PIPE_IN files. Not buildable in < 4 h (no joint harness; ATT arm ≈ hours per 4 items, MLPV md3 ≈ 4.5 h): **not started** | silicon is the first joint run; teeth in the checklist | — |
| b3 | The real hypercube (`tp_ar4_hc`, two pairwise hops) | **MISSING in sim** | every TP4 ATT sim used `tpx4_block` ideal SUM (`ATT_TP4_IMAGE_SIMS` Calculations) | **ring-h1 tooth**: the four h1 copies in the MLPV IN rings must be byte-identical per kt (main is adding it to c4_loop; mock arm exists) | 0.5 |
| b4 | 0x134D e2e ran at LANES 128; the image is LANES 256 | **PARTIAL** [M] | the world's boot stream has the L128 geometry (lane = col % 128, G12 base 12,288); the kit root is L256 (G12 base 6,144, +128 pad cols). LANES_V 256 is a packaging-stage bake | low risk: K map is LANES-independent and 0x1342 ships L256; N1–N3 netlist rows of the 0x134D record are still owed | 0.5 |
| b5 | G9 **scale** K map never exercised | **PARTIAL** | G9 scales are identical for both ranks and all quarters (`G9_scale_packed.bin` `0703b257` on r0, r1 and the TP4 files) so a wrong scale map cannot show on this checkpoint | none needed for Llama OSPREY; re-check on any checkpoint with per-group G9 scales | 0 |
| c1 | Cable plan vs the user's cables (3d c1–e1 c1, bd c1–50 c1) | **PROVEN (plan)**; silicon proof after boot | `c4_tp4.py:41-54` CABLES7 / NEW_CABLES = exactly those two; QSFP present at 3d/e1/bd/50 c1 (commit 78a4846a8). Binding is by beacon, not cage: `arm_links` requires the beacon map == CABLES7 on the 3 drafter cages of each ATT, BIND e0/e1/e2, WANT chk; `arm_mlpv_links` requires BIND 50 = 0x546, 63 = 0x475 and 3 engines up | run `c4_tp4.link_up` first thing after the cold AC | 0.2 |
| c2 | **Seating trap: MLPV rank ↔ card differs between TP2 and TP4** | **PROVEN, guarded** [M] | talon2's working TP2 kit: MLPV rank 0 = card 63, rank 1 = 50 (`f3twin kit c4_bringup.py:27-37`; stage links 17c1–63c2, 3dc0–50c3). TP4 uses the TP2 "diagonal" cables as stage links: MLPV0 = 50, MLPV1 = 63 (`c4_tp4.CARD_OF`). A TP2-style assignment boots quarters {1,3} on the card fed by ranks 0+2 | MANIFEST `card` check refuses it (tested); `arm` refuses a swapped 0x1FC word (mock RED) | 0 |
| d1 | **Identity: `c4_bringup.classify` refused B4/B1/B5** | was **BLOCKER, fixed** [M] | reg 0x100 of B4/B1/B5 = 0xE4B711B4/B1/B5 = the reg100 of the older 0xE4B712B4/12B1/12B5 rows, which come first: "tp_link4 BUILD_ID 0xe4b713b4 != label 0xe4b712b4 … id COLLISION". Found by the mock GREEN arm | `c4_bringup.py:164` now prefers the row the tp_link4 BUILD_ID names (working tree; main commits it with the Session) | done |
| d2 | MLPV 0x134D boot root | was **MISSING, done** [M] | `~/eagle4_artifacts/4c1_boot_tp4/0xE4B7134D/rank{0,1}` on box1 and talon2 (45/45 files md5 == MANIFEST on talon2). `mk_tp4_mlpv_root.py verify`: G9 1,572,864 words == the stream the GREEN xsim booted from (0 differing), G12 == xsim + zero pad, scale payloads, gamma (1,536) and GLM (393,216) stores, lm_head/embed staging equal. REDs tp2half and qswap FAIL on 393,216 words. `B.load_manifest` accepts both; refuses rank0 on card 63 and a missing verify log | — | done |
| d3 | ATT rank dirs (4 ranks, 8 heads each) | **PROVEN to boot in the mock** (junior's generator) | `…/4c1_boot_tp4/0xE4B713B{4,1,5}/rank0..3`: weights + scales + 6 D9ZERO zero-bank loads + 3 gammas + payload.ops + rank.ops (`W 1cc` 0x0/0x1/0x4/0x5). Mock boot of B4: 0x1C8 == 0, word counts == manifest. Not re-derived here: byte equality vs the AC sim's wsnap | junior (verify logs `VERIFY_TP4_*`) | — |
| d4 | 6-card arm / loop / tree | **PARTIAL: written by main, GREEN on the mock** [M] | `c4_mock7.py` (new): discover7 → link_up → boot from the real roots → arm → `c4_loop.Session` topo "7card" → tree: GREEN 3 / 0; RED 11 caught by the kit + 4 tooth-only / 0 missed | 4 tooth-only REDs = gaps in the Session (ring-h1 x3, rank bit [2] lost after arm) → main | 1 |
| d5 | 6-card golden / SGLang serve / card process (KVI at 8 heads per rank, prefill, FASTGLUE) | **MISSING** | `run_night_bench_talon2_20261110.sh:341-342` prints NOT_SUPPORTED for a bd/e1 set; `c4_sglang.py`, `c4_cardproc.py`, `c4_golden.py` had 0 TP4 lines at 07:15Z; the mock did not exercise KVI / kvonly / serve | main (c4.py / c4_sglang.py in progress) | 3–6 |
| d6 | **Correctness comparator** | **MISSING** | TP4-A8 is by design NOT bit-identical to the 4-card golden past G6: the A8 quantiser M is rank-local (`lutg_a8q.sv:12-30`; 768 elements at TP4 vs 1,536 at TP2); h1 moves ±1 LSB on ~14 % of elements (`SEVENCARD_GOLDENS_20261108.md`). Bit-identical: G4 Q/K/V, the KV rows, G5. Not: G6 partial/sum, h1, G7, r5, the whole MLPV, every tree level incl. the root. No accept or text measurement exists at TP4-A8 (review R3) | gate = digests x6, MLPV0 SoA == MLPV1, ring-h1 identity, G4 tap per rank vs the TP2 slice, repeatability; tree levels vs E2b report-only; then accept + texts. The night runner's byte-identity golden gate must not be used | 1 |
| d7 | 0x134D `EXPECT_PDI_MD5` pin; 0x134D PDI + flash adjudication | **MISSING** | routed +0.001 / +0.010 with margin; `write_device_image` running on buildxd02 at 07:4xZ; no FINAL_TRUE re-time, no marker read, no pin (`tp4_preflight_check.py` R3) | box1-two-v2 adjudicator | 1–1.5 |
| d8 | Qwen on the 6-card set | **MISSING** | images are DUAL_MODEL 1; no Qwen TP4 value arm finished (`ATT_TP4_IMAGE_SIMS` §0 "Qwen on the TP4 images: not proven"); no Qwen TP4 MLPV root (the same derivation from the qwen27b 0x1342 root is ~10 min, but has no xsim reference stream to prove against) | after Llama | 2+ |
| e1 | Depth / ABI latch | **PROVEN rule, enforced** | PERSIST latches the OUT-DMA md ABI at the FIRST AP_START: arm at the deepest depth to be used (record flow: arm d15) and run the first tree at the armed depth. `c4_tp4.arm` refuses a parked card unless `accept_latched ≥ depth`; `Session.program_cfg` refuses a first tree below the armed depth | — | 0 |
| e2 | LOOP_CTL mode bits | **PARTIAL (performance risk)** | junior's first-silicon contract is LOOP_CTL[7:4] = 0 (ITEM + COMMIT only). The 831 / 883 records ran with EAGER/EARLY_A/TREE_STOP θ −2.5. A first TP4 number without them is not comparable to the record; with them, nothing TP4 was simulated (0x134D straps EARLY_A 1, EA_EXPBURST 1, TPSAFE) | measure both; correctness first at [7:4] = 0 | — |
| e3 | K_C / FR-Spec / fold | as the record | K_C 64, VTOK 20, w2 d7 (A) / w2 d11 (B); FR-Spec has never run on 4C1 (`FRSPEC_KC_AB_4C1_20261109.md`, not adopted); G4_EMBFOLD is 0 on every LUTG image | — | 0 |
| e4 | B5 runtime columns | **host gap** | the kit drives B5 at width 0 (no `sibw` key): B5's −29 % per 2-child level is not reachable from the host yet | junior | — |

### Preflight tool
`scripts/4c1/tp4_preflight_check.py --att 0xE4B713B4 --mlpv 0xE4B7134D` (offline, no card; selftest GREEN 1 world / RED 9 caught) prints
the rows mechanically. Now: **7 PASS, 3 open (R3 PDI md5 pin, R9 6-card run path by symbol, R10 TP4 tree golden)**. R9 probes for symbols
that main's implementation does not use (it extended `Session`); read R9 as "serve path", which is still open.

## SET 2 — 4-card BATCH_MODE

| # | item | status | evidence | fix / owner | hours |
|---|---|---|---|---|---|
| f1 | MLPV 0x134C exists as a flashable image | **MISSING (close)** | buildx04 `s3LCbm2a` impl_1: Post Routing **WNS +0.014 / WHS +0.008** (margined view); no PDI at 07:4xZ; no adjudication. Arm B 0x1356: route −0.121 / WHS −0.161, still iterating | box1-two-v2: FINAL_TRUE re-time (HQ bar WHS ≥ +0.010), marker rows, BM census rows (bm_q, s1_rt_q, side FSM, r6 bank 1) | 1.5 |
| f2 | Intended ATT partner | **MISSING** | `IMAGE_B_BALANCE.md` §0: junior's merged ATT with BATCH_MODE at 0x240. **C4 0xE4B713C4 (F3t + BATCH_MODE, KV_BANKS 8) is staged, NOT launched** (queue position 2 behind B4h; `IMAGE_TABLE_7C_BATCH_20261110.md` row 15). D0 0xE4B713D0 (PF + BATCH_MODE) is in route GI1 at congestion 7. BA/BB not launched | junior: give C4 a host now | 8–10 build |
| f3 | Flashable ATTs that can pair with 0x134C today | **PROVEN only in mode 0** | F3t 0x13F9 (record, no banks) or 0x1334 (static 2 × 512 banks, non-LUTG, silicon 2-stream). With either, 0x134C is used as a 0x1342 | — | — |
| f4 | Bank-map / register agreement ATT ↔ MLPV | **PARTIAL (contract only)** | `IMAGE_A_MAXBATCH_NOLOSS.md` §0: ATT 0x240 MODE/KVB_LOG2, MLPV 0x0FC (9-bit AXI-lite), stream id 0x198[29:28], MLPV out base + s × OUT_STRIDE. STREAMCARRY (LOOP_CTL[10]) unit-gated 21/21 on 3 loop_seq bases. `j_flags` tag "NOT implemented" on the MLPV; no joint ATT+MLPV run in any mode | first C4 + 0x134C pair | — |
| g1 | Mode 0 == 0x1342 (no single-user loss) | **PROVEN (xsim)** | `IMAGE_B_MLPV_BATCHMODE.md` §4: md3 and md15, CC_CMP MATCH on 20 / 92 files, `cycles.txt` byte-identical on both ranks | — | 0 |
| g2 | Mode 2 (S1 overlap) correct | **PROVEN in the MLPV pair world only** | bit-exact vs mode 0, also under stress (S_S1_GATE engaged 4.4K / 4.7K cycles); 4 REDs fail as required; −16,109 cyc on md15 W2 (−6.7 %), 1,151 cyc per hand-off | — | — |
| g3 | Mode 2 with the REAL TAIL loop sequencer | **MISSING** | every mode-2 arm drove doorbells from a host op template ("pipelined doorbells"). On silicon the TAIL `loop_seq_d2` rings on BELL arrival and tracks `a` readiness through 0x184 / the EARLY_A count 0x1D0[7:4] (ERR 7 = "EARLY_A count ran ahead of rung items"). Mode 2 suppresses EARLY_A for handed-off children: that interaction has never been simulated | first silicon: mode 2 only after a mode-0 golden; byte-compare mode-2 trees to mode-0 trees on the same prompts; any loop ERR → back to mode 0 | 1 on silicon |
| g4 | Mode 1 (tree-grain multi-stream) on the MLPV | **MISSING** | per-stream OUT_STRIDE / stream latch have no full-drafter arm; only the ATT units (N-stream 15/15) and STREAMCARRY units ran. Full item-grain interleave needs a second loop context per card (`IMAGE_B_BALANCE.md` §4): not in any image | RTL | — |
| h1 | Who writes 0x0FC (MLPV) / 0x240 (ATT) | **MISSING** | `scripts/4c1` has no BATCH_MODE write: `c4_regmap.py:563-569` knows 0x0FC only as the ATT SIBW_RT; no row for 0x134C / 0x1356 / 0x13C4; no boot root (a relabel of the 0x1342 root is enough for 0x134C: the data is identical) | box1-two-v2: regmap rows + 0x0FC canary (0xBA7C) check + mode write while parked; relabel root + m9 scan | 1.5 |
| h2 | n-stream runner | **PROVEN on other images** | `ab_interleave_f3t_talon2_20261109.sh`, `c4_nstream.py`, `c4_hbmreplay.py` ran on F3t + 0x1342 (`INTERLEAVE_BATCHVERIFY_SILICON_20261109.md`) | — | 0 |
| h3 | Batched verify on the GPU | **PROVEN (silicon), with a caveat** | BV4 + HBMREPLAY **730.25 / 716.99 tok/s @ 5.90** vs serial 665.33 / 665.50 on the same launcher (+8.7 %). Batched verify changes 6–8 / 80 texts vs bs 1 (GPU-only does the same): a batched-forward numerics class, flagged for any product use | — | 0 |
| h4 | What number to expect, against what | **defined** | single stream: F3t 831.14 @ 5.89 (humaneval 80×256 w2 d7) / 883.1 serve config. 2-stream aggregate: 0x1334 KVB2 638.84 / 655.12 (below its own single-stream 758). Best measured multi-request: 730 (BV4 + HBMREPLAY on F3t). 0x134C mode 2 can add at most the MLPV's exposed share: MLPV-only window 219–273 µs of a 2,358 µs tree [M] × 6.7 % → **≈ +0.6–0.8 % single stream** [C]; a BATCH_MODE pair's aggregate ceiling is the card window: 424 trees/s tree-grain [C, `IMAGE_B_BALANCE.md`] | — | — |

## Ordered pre-flash checklist

**SET 1 (do in this order; each line names its proof):**
1. 0x134D PDI written → FINAL_TRUE re-time on a dcp copy (≥ +0.030 / ≥ +0.010) → marker FLASH-RELEVANT rows → N1–N3 netlist rows
   (ring-B adder, hop_xport2, SOLO engines) → `preflash_gate.sh` dry run → pin `EXPECT_PDI_MD5[0xE4B7134D]`.
2. `python3 -I scripts/4c1/tp4_preflight_check.py --att 0xE4B713B4 --mlpv 0xE4B7134D` → R1–R8 PASS.
3. `python3 -I scripts/4c1/m9_4kb_scan.py ~/eagle4_artifacts/4c1_boot_tp4` on the host that boots (talon2) → PASS on all six rank dirs.
4. `python c4_mock7.py --boot` with the four verify logs in `EAGLE4_4C1_VERIFY_LOG` → GREEN, and the 4 tooth-only REDs flipped to
   kit-caught once the Session carries the ring-h1 and 0x1CC re-read teeth.
5. Flash 50/63 (-p 0, detached, FLASH_EXIT=0 ×2) → cold AC → `ami_tool overview` READY ×7 → QDMA.
6. Identity two-factor (reg100 + tp_link4 BUILD_ID + md5) with the patched classify. Card 99: read only.
7. `c4_tp4.link_up`: beacon map == CABLES7 (the two new cables), BIND on all 4 ATTs and both MLPVs, ENG0/1/2 up.
8. Boot with `EAGLE4_4C1_BOOT_ROOT=~/eagle4_artifacts/4c1_boot_tp4`: rank0 dir → card 50, rank1 → 63 (TP4 seating); 0x1C8 == 0 ×6;
   0x74 never reads 0x4 (timeout); G8 lm-head duplicate staged on both MLPVs.
9. Arm at the DEEPEST depth to be used (d15), LOOP_CTL[7:4] = 0; TAIL_IN ring A for ranks 0/1, ring B for 2/3 with read-back; MLPV loop
   SOLO_STAT 0x5010 / TPSAFE 0xE7; 0x1FC a0ff0001 on 50, a1ff0101 on 63; 0x1CC 0x0 / 0x1 / 0x4 / 0x5.
10. First tree at the armed depth. Hard teeth: no loop ERR (6 / 9 name the join), digests equal on 4 HEADs + 2 TAILs, MLPV0 SoA == MLPV1,
    the four ring h1 copies byte-identical and non-zero for every kt, same tree twice = identical SoA, G4 tap per rank == the TP2 slice.
    Tree levels vs E2b: report only.
11. Serve: 0 discarded trees, accept and texts vs the F3t record on the fixed benchmark; then the speed levers (EAGER / EARLY_A /
    TREE_STOP) one at a time, each with the step-10 teeth.
12. Demo guard unchanged: no golden pass + measured win over 831.14 / 883.1 by 14:00Z → restore F3t 0x13F9 + 0x1342.

**SET 2:**
1. Adjudicate 0x134C (FINAL_TRUE, hold bar, BM census). 2. Host: regmap row, relabelled 0x1342 boot root + m9 scan, 0x0FC canary read.
3. Flash 0x134C on 50/63 next to F3t; golden in mode 0 must be byte-identical to the F3t + 0x1342 golden (this one IS the right comparator).
4. Mode 2 only while parked, then mode-2 trees byte-compared to mode-0 trees; a loop ERR 7 / 9 ends the experiment.
5. BATCH_MODE banks need C4 (launch it) or D0; until then multi-request = BV4 + HBMREPLAY on F3t (730 measured).

## What was created (all create-only)
- `scripts/4c1/m9_4kb_scan.py` — the 4 KB burst gate.
- `scripts/4c1/mk_tp4_mlpv_root.py` — 0x134D rank-dir generator + independent verifier + RED arms.
- `scripts/4c1/tp4_preflight_check.py` — offline readiness rows.
- `scripts/4c1/c4_mock7.py` — 6-card TP4 mock world + self-test over the real Session.
- Boot root `~/eagle4_artifacts/4c1_boot_tp4/0xE4B7134D/rank{0,1}` + `verify_tp4_mlpv_0xE4B7134D_20261110.log` (box1, talon2).
- One line changed in `scripts/4c1/c4_bringup.py` (classify, reg100 collision), left uncommitted for main's Session commit.

## Calculations and hypothesis
- [C] 0x134C mode-2 single-stream gain ≈ (273.5 µs MLPV-only exposure / 2,357.7 µs tree window) × 6.7 % ≈ 0.8 %; with the 219.4 µs figure
  ≈ 0.6 %. The xsim −6.7 % is MLPV cycles per tree, not tok/s.
- [C] The ring-h1 tooth detects a wrong hypercube only if the wrong sum differs between ranks; a hypercube that gives all four ranks the
  SAME wrong sum passes it. The TP4-A8 bit-true h1 golden (`tp4_card_goldens_a8.py`, stimulus stim_md7) closes that when the first tree
  replays that stimulus.
- [C] `c4_mock7` encodes one reading of the SOLO join (ERR 4 / 6 / 9 from unequal ARM_TREE / ARM_KT0 / root header). A shared misreading
  between the mock and the Session passes both; the link gate `tb_solo` is the RTL evidence for the join itself.
