# ATT-combo silicon window 2026-11-16: COMBO3 0xE4B713D6 + D5F 0xE4B713DB on talon2

Program 1: the 7-card Llama-3.2-3B drafter (ATT TP4 x4 on 17/3d/bd/e1 + MLPV M1361 e24deb3f x2 on 50/63, MLPV not flashed).
Window agent: box1-two-v2 (lock OWNER `box1-two-v2-attcombowin`). Runbook `WINDOW_RUNBOOK_ATTCOMBO_20261116.md` (md5 e179f721), adjudication `ADJUDICATION_0x13D6_0x13DB_20261116.md` (063938e8f).
Window dir talon2 `~/eagle4_talon2_4c1/att_combo_20261116/` (MD5SUMS.txt -c clean 13:29Z); run dir `~/eagle4_talon2_4c1/runs/acwin_20261116/{d6,db,restore}`.
Bars (same day, record set ATT 0x13BE a21ca534 x4 + M1361 x2): W2 938.94 @6.91 texts 1b0779cf; W4 TREE_STOP=0 756.05 @7.20 texts 54f88d49.

## Log (UTC, evidence = talon2 run-dir files)
- 13:29:18 lock TAKEN after box1-two-v2-136Bwin FREE 13:26:49 (record set, boot b4f82817, fp 51524c4f). MLPV = M1361.
- 13:32:02 COMBO3 flash launched on 17/3d/bd/e1 -p 0 detached (PDI d2e737d7, 79,827,088 B, FPT 16a5f792, AMC 32; record 6d7cfa2d; pre-flash ATT 0x1D4 0x3000 x4, ami READY 7). Logs `tp4_7c_20261110/flash_TP4D6_{17,3d,bd,e1}_20261116_133202.log`.
- W4 TREE_STOP=0 driver: create-only copies `ac_serve_w_ts.sh` (1c127f85) / `ac_steps_ts.sh` (5f433a64) = the window scripts with ONLY the record MENV `EAGLE4_PAIR_TREE_STOP=${TSTOP:-1}` (the runbook's ac_serve_w.sh hard-codes TREE_STOP=1 under SMODES=record; the 756.05 bar is TS=0, W4_SERVE_BOUND_20261115.md:203-211).
- 14:09:35 FLASH_EXIT=0 x4 (complete 1 each). 14:09:48 PDU cold AC #1 (box1 `/var/tmp/box1two_scratch_20261116/coldac1_attcombowin.log`): PDU 14:08:38 -> OFF 14:12 / ON 14:18, shutdown -h 14:11:30, post-OFF/ON reads empty, backup manual ON (attempt 2: outlet8 447 W, outlet24 415 W), boot 14:22:21 = **79cc9d1b**, COLDAC_SCRIPT_EXIT=0.
- 14:24 COMBO3 bring-up: k149 insmod rc 0 x2, ami READY 7, QDMA x6 rc 0 Roundtrip PASS x6, c4 status pre-QDMA rc 0 + post-QDMA pre-arm rc 0, **IDENT_OK 6/6** (ATT reg100 0xe4b711d6 / tp 0xe4b713d6 x4, MLPV 0xe4b71161 / 0xe4b71361 x2). kvread prearm **CLEAN 4/4** (0x228 0x08000001 legacy, 0x200 0x4d430000, 0x028 0, 0x1D4 0x3000).
- 14:25:12 arm + first + record golden: golden rc 50 / gate rc 0 (vacuous by itself), **GOLDEN_FP PASS fp=51524c4f** (12 results, golden_record_142512.json f1340236), **EXACT_TOOTH PASS x2** (record + first).

### COMBO3 0xE4B713D6 serve rows (boot 79cc9d1b; d11 fa3 fg4 cp1 K_C 64 humaneval 80x256, record levers HX_MMIO7/O7/S7 v; logs talon2 `runs/acwin_20261116/d6/serve/`)
| row | W | TREE_STOP | KV1H2C | tok/s | accept | steps | rc / errs | texts (80) | 0x1D4 after |
|---|---|---|---|---|---|---|---|---|---|
| s15 smoke (d15) | 2 | 1 | on | 959.34 | 7.17 | 2735 | 0 / 0 | 3c26f0b0 | 0x3000 x4 (0x228 0x00000001 AUTO) |
| t1 | 2 | 1 | on | 960.23 | 6.91 | 2839 | 0 / 0 | 1b0779cf = record | 0x3000 x4 |
| t2 | 2 | 1 | on | 960.97 | 6.91 | 2839 | 0 / 0 | 1b0779cf = record | 0x3000 x4 |
| t3 | 2 | 1 | on | 963.10 | 6.91 | 2839 | 0 / 0 | 1b0779cf = record | 0x3000 x4 |
| off1 | 2 | 1 | **off** (kv1h2c_lines 0) | 955.22 | 6.91 | 2839 | 0 / 0 | 1b0779cf = record | 0x3000 x4 |
| w4a | 4 | **0** | on | 792.67 | 7.20 | 2715 | 0 / 0 | 54f88d49 = record W4 | 0x3000 x4 |
| w4b | 4 | **0** | on | 792.32 | 7.20 | 2715 | 0 / 0 | 54f88d49 = record W4 | 0x3000 x4 |

- **W2 mean 961.43 (3 rounds) vs bar 938.94 = +2.40 %**, texts == record x3, accept 6.91 (same as record). KV1H2C on-vs-off: 961.43 vs 955.22 (1 off round) = +0.65 %.
- **W4 TS=0 mean 792.50 vs 756.05 = +4.82 %**, texts 54f88d49 x2 = record W4, accept 7.20. TS=0 verified in the logs: `EAGLE4_PAIR_TREE_STOP=0`, `[4c1-ts] PAD` lines 0.
- Zero KV-reader aborts on COMBO3 (0x1D4 0x00003000 x4 after all 7 rounds). **COMBO3 meets every keep-rule check.**
- 14:36:16 D5F flash launched on 17/3d/bd/e1 -p 0 detached (PDI 24849a8f; record 6fbc36a3; pre-flash ATT 0x1D4 0x3000 x4, ami READY 7). Logs `tp4_7c_20261110/flash_TP4DB_{17,3d,bd,e1}_20261116_143616.log`. Until cold AC #2 the cards keep running COMBO3 (boot 79cc9d1b).
- 15:14:37 FLASH_EXIT=0 x4. 15:14:49 PDU cold AC #2 (box1 `/var/tmp/box1two_scratch_20261116/coldac2_attcombowin.log`): PDU 15:13:35 -> OFF 15:17 / ON 15:23, shutdown -h 15:16:25, post-OFF/ON reads empty, backup manual ON attempt 1 (outlet8 316 W, outlet24 397 W 15:25:40), boot 15:27:13 = **d4ec5f32**, COLDAC_SCRIPT_EXIT=0.
- 15:29 D5F bring-up: insmod rc 0 x2, ami READY 7, QDMA x6 rc 0 PASS, c4 status x2 rc 0, **IDENT_OK 6/6** (ATT 0xe4b711db / 0xe4b713db x4, MLPV 0xe4b71161 / 0xe4b71361), kvread prearm CLEAN 4/4 (0x228 = 0 STRAPLESS). 15:29:47 arm OK depth 15, golden rc 50 / gate 0 x2, **GOLDEN_FP PASS 51524c4f**, **EXACT_TOOTH PASS x2**.
- **dsh-probe deadlock (15:31-15:38) [M]**: the 15:30 dsh-probe card_run (not mine; pid 4257, sshd 4247, started 15:30:05) got `lock/card.lock` at 15:31:18, in the 0-second gap between my smoke steps. It then held that flock while it polled `~/TALON2_CARD_LOCK` for a non-OWNER last line (card_run.sh:15-17). The window serve's card_run (pid 4814, private T2LOCK=steps.lock) was blocked on the same flock for 7 min. Nothing was killed. 15:38:43 I appended a `NOTE ... window STILL HELD (not FREE)` line. The probe ran its read-only c4 status (OWNER/FREE 15:38:48, rc 0) and the serve resumed at 15:39:55. 15:42:23 I re-took the OWNER line (`PREV_OWNER=dsh-probe ac_lock.sh take`). The t1 row's wall time includes the stall, but its tok/s does not (the engine had not started).
- Box1 ssh for the W4 phase dropped (`client_loop: send disconnect: Broken pipe`, cloudflared) after w4b started. The talon2 phase still completed: SUMMARY rows w4a/w4b, `ac_kvread_post_serve_154450.json`, no NO_REARM.

### D5F 0xE4B713DB serve rows (boot d4ec5f32; same settings; KV1H2C n/a, 0x228 = 0)
| row | W | TREE_STOP | tok/s | accept | steps | rc / errs | texts (80) | 0x1D4 after |
|---|---|---|---|---|---|---|---|---|
| s15 smoke (d15) | 2 | 1 | 942.49 | 7.17 | 2735 | 0 / 0 | 3c26f0b0 (= COMBO3 smoke) | 0x3000 x4 |
| t1 | 2 | 1 | 942.14 | 6.91 | 2839 | 0 / 0 | 1b0779cf = record | 0x3000 x4 |
| t2 | 2 | 1 | 940.34 | 6.91 | 2839 | 0 / 0 | 1b0779cf = record | 0x3000 x4 |
| t3 | 2 | 1 | 945.12 | 6.91 | 2839 | 0 / 0 | 1b0779cf = record | 0x3000 x4 |
| w4a | 4 | 0 | 782.18 | 7.20 | 2715 | 0 / 0 | 54f88d49 = record W4 | 0x3000 x4 |
| w4b | 4 | 0 | 782.26 | 7.20 | 2715 | 0 / 0 | 54f88d49 = record W4 | 0x3000 x4 |

- **D5F W2 mean 942.53 vs 938.94 = +0.38 %** (inside the record's 937.25-942.20 round spread). **W4 TS=0 782.22 vs 756.05 = +3.46 %.** Zero KV aborts.
- **COMBO3 vs D5F on the same day: W2 961.43 vs 942.53 (+2.0 %), W4 792.50 vs 782.22 (+1.3 %).** D5F lacks HCUT and KV1H2C. The KV1H2C-off COMBO3 round (955.22) still beats every D5F round.

## Keep decision [D] (runbook section 4)
- Both images pass every check: fp 51524c4f, tooth x2, d15 smoke clean, 0x1D4 0x3000 after every round, rc/errs 0, W2 texts 1b0779cf, W2 mean > 938.94.
- The higher mean wins: **keep COMBO3 0xE4B713D6 d2e737d7 on ATT x4** (new W2 record 961.43 @6.91).
- 15:52:23 COMBO3 keeper re-flash launched, `RUNB=runs/acwin_20261116b` (record 6d7cfa2d, PDI OK d2e737d7). Logs `tp4_7c_20261110/flash_TP4D6_{17,3d,bd,e1}_20261116_155223.log`.
- Then: cold AC #3 -> STAGE=d6 RUNB=acwin_20261116b drv/qdma/status2/ident/kvread/bringup -> fp + tooth -> one W2 confirm round (KV1H2C on) -> FREE.
- 16:29:35 keeper FLASH_EXIT=0 x4. The coordinator asked at 16:20Z to skip the restore flash and FREE on D5F; that order was superseded by the 16:2xZ plan change ("run the normal tail"), and the flash, already ~75 % done, was never stopped. 16:29:54 PDU cold AC #3 (box1 `/var/tmp/box1two_scratch_20261116/coldac3_attcombowin.log`): PDU 16:28:42 -> OFF 16:32 / ON 16:38, scheduled OFF read 0 W, scheduled ON read On 13 W / 12 W (no manual ON), boot 16:42:13 = **321f3241**, COLDAC_SCRIPT_EXIT=0.
- 16:44-16:45 keeper bring-up (RUNB `runs/acwin_20261116b`): insmod rc 0, ami READY 7, QDMA x6 PASS, c4 status x2 rc 0, **IDENT_OK 6/6** (0xe4b711d6 / 0xe4b713d6 x4, M1361 x2), kvread CLEAN 4/4, arm OK depth 15, **GOLDEN_FP PASS 51524c4f**, **EXACT_TOOTH PASS x2**. Restore-confirm rc1 W2 d11 **956.62 @6.91**, texts 1b0779cf.
- 16:44 lock: the coordinator relayed that box1-qwen-junior released talon2 to box1-two-v2; OWNER re-asserted 16:44:35 for the user-ordered 7-card depth x width sweep. 16:47 census + 16:48-17:50 sweep + K_C cells + VTIMING, 0 aborts: **`DW_SWEEP_7CARD_COMBO3_20261116.md`**.
- cloudflared drops (box1 ssh `Broken pipe`) twice more during phases: the talon2-side phases completed every time. From 16:54 the sweep driver launches each phase detached ON talon2 (setsid, log in `runs/acwin_20261116b/dw_sweep/logs/`). The box1 probe watcher (`dw_sweep_combo3_20261116/probewatch_box1.sh`) never fired: no dsh-probe got stuck at 17:00 or 17:30.
- **17:52:00Z FREE**: COMBO3 0xE4B713D6 d2e737d7 x4 + M1361 e24deb3f x2 RUNNING and STORED, boot 321f3241, fp 51524c4f, armed. Next: box1-qwen-junior Qwen sweep (139A + 139D).

## Window verdict
- **New 7-card W2 record: COMBO3 0xE4B713D6 + M1361 = 961.43 @6.91 (t1-t3, boot 79cc9d1b), +2.40 % vs 938.94, texts == record.** It is kept on ATT x4. Every W2 d11 K_C 64 round on COMBO3 today: 960.23 / 960.97 / 963.10 / 956.62 / 957.42 / 962.78 / 961.06 (vt) = 960.31 over 7.
- D5F 0xE4B713DB: correct, 942.53 (+0.38 %), not kept. The KVRD credit fix is not needed on COMBO3's 0-shared-link placement: 0 KV aborts over 3 boots and ~60 serve rounds, d15 included.
- Cold ACs: 3 (D6, DB, keeper), all through the PDU with outlets 3-24 back ON.
