# 0xE4B71361 (MLPV_G8_OWN) on talon2 silicon, 2026-11-12 (window agent for box1-two-v2)

Program 1: the 7-card Llama-3.2-3B drafter (ATT TP4 ranks 17/3d/bd/e1 + MLPV 50/63; card 99 never opened). Motto: engineer toward the fastest flashable PDI.
Lock: `OWNER box1-two-v2-g8own 2026-11-12T16:49:03Z` (taken by box1-two-v2 on junior's `FREE 16:48:11Z`).
Evidence: `g8own_silicon_20261112/` (`ship/` = every script shipped to talon2 `~/eagle4_talon2_4c1/g8own_20261112/`, `box1/` = console outputs,
`talon2/` = run-dir copies). talon2 run dir: `~/eagle4_talon2_4c1/runs/g8own_1361_20261112/` (restore boot: `.../restore/`).

## Adjudication note (coordinator, 16:5xZ)
- 0x1361 true setup **+0.000177 ns is below the +0.030 FINAL_TRUE release bar.** Flashing it is a **deliberate exception on the 0x1360 precedent**
  (record MLPV, adjudicated GO 11-11 at +0.001 / +0.010 true, same no-margin-hook lineage). Hold +0.010 meets its bar.
- **Condition: any silicon anomaly (hang, watchdog abort, text mismatch, gate FAIL) = quarantine 0x1361 and restore 0x1360 in this window.**
- One party drives the PDU: this window agent (junior gets a one-line heads-up first).

(results below)

## RESULT (box1-two-v2, 2026-11-12 18:02Z): 0x1361 = NEW 7-CARD RECORD, correct; kept on the cards (no restore)
| serve (UTC) | tok/s | accept | steps / tokens | errs | c4_hx7 | texts |
|---|---|---|---|---|---|---|
| g1 17:56:02-17:57:18 | 904.97 | 6.91 | 2839 / 19615 | 0 | 1 | 1b0779cf (80) = record |
| g2 17:57:18-17:58:26 | 902.40 | 6.91 | 2839 / 19615 | 0 | 1 | 1b0779cf (80) = record |
| g3 17:58:26-17:59:32 | 900.93 | 6.91 | 2839 / 19615 | 0 | 1 | 1b0779cf (80) = record |
- **Configuration:** 0x13BC 41e69bde ×4 (17/3d/bd/e1) + 0xE4B71361 626f8c0c ×2 (50/63); kit kit7c_52ee0696_g8own (= the hx7 kit + 3 c4_regmap rows); d11 fa3 w2 FG4 CP1 K_C 64; record modes; humaneval 80×256; XF_EXTRA `EAGLE4_4C1_HX_MMIO7=1 HX_MQ=0 HP_CAPHOST=1 HP_IMGBYTES=1` (the same env as junior's 889.26 run).
- **Logs:** talon2 `~/eagle4_talon2_4c1/runs/g8own_1361_20261112/serve/t4w_hx7_kc64_g{1,2,3}_*.log`; driver output in `phase_serve1.out`.
- **Mean 902.77 tok/s.** That is +1.52 % vs the same-day record-set confirm 889.26 (0x13BC+0x1360), and +1.40 % vs the 4-run record mean 890.30. The slowest 0x1361 run (900.93) is faster than the best 0x1360 run (891.15). The model predicted +1.2…+2.1 % (DESIGN_FABLE_MLPV_CHAIN_7C §3 lever 1).
- **M1:** kc_cfg 0x150 = 0x40 on both MLPV cards (EQUAL), read after both goldens and after every serve. **KV reader:** ATT 0x1D4 = 0x00003000 ×4 after every serve; no abort. **Goldens:** rc 50 / gate rc 0, read directly. The 12 level tables are identical to the 0x13BC+0x1360 golden af932f7b.
- **Decision (box1-two-v2):** no anomaly occurred, so the quarantine condition never fired. 0x1361 stays on 50/63 as the new record MLPV, and the restore is skipped (saving one cold AC).
  - The +0.000177 setup exception stands. A closure arm (buildxd00) and the G8OWN63 BalancedSLR hedge (buildzd04) are running toward a higher-margin re-flash candidate. G8_OWN v2 is in design on the 0x1360 base.
- **Parent baseline:** no 0x13BC+0x134D serve exists. The nearest are B4+0x134D plain 859.06 and B4+0x1360 hx7 879.57.
- Flash 16:54:03Z → FLASH_EXIT=0 ×2 17:36:50Z. Cold AC #1: halt 17:39:38, OFF 17:41:54, ON 17:47:51, boot 17:50:01, 7 READY 17:51:38. Boot_id 3f89b99e.

## Census on the record set (talon2 18:00-18:01Z, c4 bench, kit kit7c_52ee0696_g8own, 100 trees each, ctx 566)
| arm | issue_done p50 ms | trees/s | items/tree | us/level | vs 0x13BC+0x1360 TP4 census |
|---|---|---|---|---|---|
| md11 w2 | 2.959 | 338.9 | 21 | **269** | 280.0 → 269 (−3.9 %) |
| md7 w2 | 1.963 | 498.5 | 13 | 280 | — |
- Every bench run: BENCH_EXIT=0, 0 trees discarded, 0 differing from tree 0.
- Evidence: talon2 `runs/g8own_1361_20261112/census/bench_P{11,7}w2_*.json`.

## M1361 (0xE4B71361 e24deb3f, re-flash 2026-11-13) on talon2: verdict KEEP (box1-two-v2 window agent, 2026-11-13 18:55Z)
Boot 18:30:05 (boot_id c4541757, kernel 6.8.0-149; card drivers hand-insmod'ed from `~/drv_k149_m1361_20261113/`). Cards: 0x13BC ×4 (17/3d/bd/e1) + M1361 e24deb3f ×2 (50/63). Lock `OWNER box1-two-v2-m1361 17:29:25Z`.
| serve (UTC) | tok/s | accept | steps / tokens | errs | c4_hx7 | hxo7 stock / late_fail | hxs7 v_fail | texts |
|---|---|---|---|---|---|---|---|---|
| g1 18:39:15-18:50:43 | 930.34 | 6.91 | 2839 / 19615 | 0 | 1 | 0 / 0 | 0 | 1b0779cf (80) |
| g2 18:50:43-18:51:52 | 934.64 | 6.91 | 2839 / 19615 | 0 | 1 | 0 / 0 | 0 | 1b0779cf (80) |
| g3 18:51:52-18:52:57 | 930.91 | 6.91 | 2839 / 19615 | 0 | 1 | 0 / 0 | 0 | 1b0779cf (80) |
- Kit `kit7c_52ee0696_g8own_hxs7_m1361`; XF_EXTRA `EAGLE4_4C1_HX_MMIO7=1 EAGLE4_4C1_HX_MQ=0 EAGLE4_4C1_HP_CAPHOST=1 EAGLE4_4C1_HP_IMGBYTES=1 EAGLE4_4C1_HX_O7=1 EAGLE4_4C1_HX_S7=1 EAGLE4_4C1_HX_S7_PARTS=v`; d11 fa3 w2 FG4 CP1 K_C 64; record modes; humaneval 80×256. Driver `m1361_20261113/m1361_serve_w.sh` via `m1361_steps.sh serve`; rows in talon2 `runs/m1361_20261113/serve/SUMMARY.tsv`, `SUMX.tsv`, `phase_serve1.out`.
- Mean **931.96 tok/s**. The record band is 926–934 @ 6.91. g1 includes the first-after-boot target prefill CUDA-graph capture of 561.73 s (later rounds: 5.57 s), which does not enter the timed window.
- After every round: M1 0x150 = 0x40 on both MLPV cards (EQUAL), and ATT 0x1D4 = 0x00003000 ×4.
- **Census md11 w2** (18:54:53Z, `runs/m1361_20261113/census/bench_P11w2_md11_w2.out`): BENCH_EXIT=0; 100 trees ok, 0 discarded, 0 differ. issue_done p50 2.878 ms; 345.1 trees/s; **262 µs/level** (11-12 0x1361: 269; 0x1360: 280.0).
- **Verdict: KEEP M1361 e24deb3f on 50/63.** Texts are identical, tok/s is inside the record band, and census µs/level is not slower. No restore is needed: the MLPV to restore is M1361 e24deb3f.
- Followed by the user benchmark suite on the same boot: `reports/results/BENCH_FPGA_VS_GPU_20261113.md`. All 10 FPGA serves there (5 datasets × 2) ran rc 0, errs 0, with M1 0x150 = 0x40 EQUAL after each.
- Last on the boot: junior's TP4 **W8** comparator `w8a`, which **aborted** at 19:59:28Z with rc 247 (`TreeDiscarded: MLPV0@50:00.0: STALE TREE READBACK`, poison survived at region0[0] / region4[647]). It left MLPV 0x1D4 = 0x00009bff and 0x1E8 = 0x00149bff, so the cards need a cold AC before any arm. This is a W8-path fault, not a W2 one; it does not change the KEEP verdict, which rests on the W2 record config. FREE 20:02:22Z.
