# Strata bench summary

*Generated 2026-10-05T06:05:48Z by `harness/summary.py` from `/workshop/bench-archive/plans-2026-10-02/strata-bench/records`; no figure is typed by hand. Internal: it names the laptop. PREREG.md sha256 `withheld-2026-09-29`, registered 2026-10-02T19:26:17Z, gate block verbatim: True.*

Strata (github.com/Niko1221/Strata) is an unrelated project; strata2signal has no affiliation with it.

## 1 · Every attempt and its outcome

| arm | phase | attempt | marker | sentence | source |
|---|---|---:|---|---|---|
| L-IQ3 | fit | a1 | STOPPED | STOPPED: memory pressure: PSI memory full avg10 11.52 % >= 10 % | `L-IQ3/fit/a1/STOPPED.json` |
| L-IQ3 | fit | a2 | REFUSED | REFUSED (G0): devices []: one CUDA device and no Vulkan device is the rule; VOID: the runner log does not say how many layers it offloaded | `L-IQ3/fit/a2/REFUSED.json` |
| L-IQ3 | fit | a3 | STOPPED | STOPPED: memory pressure: PSI memory full avg10 12.08 % >= 10 % | `L-IQ3/fit/a3/STOPPED.json` |
| L-IQ3 | fit | a4 | STOPPED | STOPPED: memory pressure: PSI memory full avg10 12.4 % >= 10 % | `L-IQ3/fit/a4/STOPPED.json` |
| L-IQ3 | fit | a5 | REFUSED | REFUSED (G0): /health never answered ({'ready': False, 'exited': 1, 'seconds': 37.19}) | `L-IQ3/fit/a5/REFUSED.json` |
| L-IQ3 | fit | a6 | RESULT | the phase ran to its end | `L-IQ3/fit/a6/RESULT.json` |
| L-IQ3 | fit | a7 | RESULT | the phase ran to its end | `L-IQ3/fit/a7/RESULT.json` |
| L-IQ3 | fit | a8 | REFUSED | REFUSED (G0): /health never answered ({'ready': False, 'exited': 1, 'seconds': 45.05}) | `L-IQ3/fit/a8/REFUSED.json` |
| L-IQ3 | fit | a9 | REFUSED | REFUSED (G0): /health never answered ({'ready': False, 'exited': 1, 'seconds': 86.04}) | `L-IQ3/fit/a9/REFUSED.json` |
| L-IQ3 | quality | a1 | STOPPED | STOPPED: the inner runner raised: SeatError: HTTP None: RemoteDisconnected: Remote end closed connection without response | `L-IQ3/quality/a1/STOPPED.json` |
| L-IQ3 | quality | a2 | STOPPED | STOPPED: the inner runner raised: SeatError: HTTP None: RemoteDisconnected: Remote end closed connection without response | `L-IQ3/quality/a2/STOPPED.json` |
| L-IQ3 | speed | a1 | RESULT | the phase ran to its end | `L-IQ3/speed/a1/RESULT.json` |
| L-IQ3-t8 | speed | a1 | RESULT | the phase ran to its end | `L-IQ3-t8/speed/a1/RESULT.json` |
| L-IQ3-fa0 (fa-off) | load | a1 | STOPPED | STOPPED (exited_before_ready; re-open with --attempt N+1): the server exited before /health answered and its log carries no line G0 refuses on (no Vulkan device | `L-IQ3-fa0/load/a1/STOPPED.json` |
| L-IQ3-fa0 (fa-off) | load | a2 | RESULT | the phase ran to its end | `L-IQ3-fa0/load/a2/RESULT.json` |
| L-IQ3-fa0 (fa-off) | quality | a1 | RESULT | the phase ran to its end | `L-IQ3-fa0/quality/a1/RESULT.json` |
| S-IQ3 | load | a1 | RESULT | the phase ran to its end | `S-IQ3/load/a1/RESULT.json` |
| S-IQ3 | quality | a1 | RESULT | the phase ran to its end | `S-IQ3/quality/a1/RESULT.json` |
| S-IQ3 | speed | a1 | RESULT | the phase ran to its end | `S-IQ3/speed/a1/RESULT.json` |
| S-IQ3 | speed | a2 | STOPPED | STOPPED: memory pressure: PSI memory full avg10 10.65 % >= 10 % (steady; the box's = the larger of system-wide and user@) [PSI full avg10: engine unit's own 0 % | `S-IQ3/speed/a2/STOPPED.json` |
| S-IQ3-noMTP | load | a1 | STOPPED | STOPPED (exited_before_ready; re-open with --attempt N+1): the server exited before /health answered and its log carries no line G0 refuses on (no CPU refusal,  | `S-IQ3-noMTP/load/a1/STOPPED.json` |
| S-Q2 | load | a1 | RESULT | the phase ran to its end | `S-Q2/load/a1/RESULT.json` |
| S-Q2 | quality | a1 | RESULT | the phase ran to its end | `S-Q2/quality/a1/RESULT.json` |
| S-Q2 | speed | a1 | RESULT | the phase ran to its end | `S-Q2/speed/a1/RESULT.json` |
| S-Q2 | speed | a2 | RESULT | the phase ran to its end | `S-Q2/speed/a2/RESULT.json` |
| L-Q2 | fit | a1 | RESULT | the phase ran to its end | `L-Q2/fit/a1/RESULT.json` |
| L-Q2 | fit | a2 | RESULT | the phase ran to its end | `L-Q2/fit/a2/RESULT.json` |
| L-Q2 | fit | a3 | RESULT | the phase ran to its end | `L-Q2/fit/a3/RESULT.json` |
| L-Q2 | fit | a4 | RESULT | the phase ran to its end | `L-Q2/fit/a4/RESULT.json` |
| L-Q2 | fit | a5 | RESULT | the phase ran to its end | `L-Q2/fit/a5/RESULT.json` |
| L-Q2 | quality | a1 | STOPPED | STOPPED: memory pressure: PSI memory full avg10 19.58 % >= 10 % (steady; the box's = the larger of system-wide and user@) [PSI full avg10: engine unit's own 41. | `L-Q2/quality/a1/STOPPED.json` |
| L-Q2 | speed | a1 | RESULT | the phase ran to its end | `L-Q2/speed/a1/RESULT.json` |
| L-Q2 | speed | a2 | RESULT | the phase ran to its end | `L-Q2/speed/a2/RESULT.json` |
| L-Q2-fa0 (fa-off) | load | a1 | STOPPED | STOPPED (exited_before_ready; re-open with --attempt N+1): the server exited before /health answered and its log carries no line G0 refuses on (no Vulkan device | `L-Q2-fa0/load/a1/STOPPED.json` |
| L-Q2-fa0 (fa-off) | load | a2 | RESULT | the phase ran to its end | `L-Q2-fa0/load/a2/RESULT.json` |
| L-Q2-fa0 (fa-off) | quality | a1 | RESULT | the phase ran to its end | `L-Q2-fa0/quality/a1/RESULT.json` |
| S-HEAD | load | a1 | RESULT | the phase ran to its end | `S-HEAD/load/a1/RESULT.json` |
| S-HEAD | quality | a1 | RESULT | the phase ran to its end | `S-HEAD/quality/a1/RESULT.json` |
| S-HEAD | speed | a1 | RESULT | the phase ran to its end | `S-HEAD/speed/a1/RESULT.json` |
| S-HEAD | speed | a2 | STOPPED | STOPPED: memory pressure: PSI memory full avg10 11.98 % >= 10 % (steady; the box's = the larger of system-wide and user@) [PSI full avg10: engine unit's own 0.0 | `S-HEAD/speed/a2/STOPPED.json` |
| L-Q2 | speed | a1 | preflight (n-cpu-moe) | no FROZEN-NCMOE-L-Q2.json: run `driver.py fit` then `driver.py freeze-ncmoe` first | `L-Q2/speed/PREFLIGHT-REFUSALS.jsonl` |
| S-Q2 | speed | a2 | quiet check (PASS) | 2026-10-03T05:03:46Z–05:09:46Z: quiet for 60 s: no busy unit or process, the 1-minute load average below 2 at all 13 polls (max 1.93) | `S-Q2/speed/QUIET-CHECK.jsonl` |
| S-IQ3 | speed | a2 | quiet check (REFUSED) | 2026-10-03T05:13:26Z–05:23:27Z: REFUSED quiet-check: not quiet for 60 s straight within 600 s; now: ovv-audit-a-gates-p3-052247.scope (rule 4: a lane scope); load1 2.88 (line 2); las | `S-IQ3/speed/QUIET-CHECK.jsonl` |
| S-IQ3 | speed | a2 | quiet check (REFUSED) | 2026-10-03T05:47:03Z–05:57:04Z: REFUSED quiet-check: not quiet for 60 s straight within 600 s; now: gao-audit-attack-4.scope (rule 4: a lane scope); load1 2.34 (line 2); last busy: g | `S-IQ3/speed/QUIET-CHECK.jsonl` |
| S-IQ3 | speed | a2 | quiet check (PASS) | 2026-10-03T07:23:53Z–07:24:53Z: quiet for 60 s: no busy unit or process, the 1-minute load average below 2 at all 13 polls (max 0.29) | `S-IQ3/speed/QUIET-CHECK.jsonl` |
| S-HEAD | speed | a2 | quiet check (REFUSED) | 2026-10-03T05:23:27Z–05:33:28Z: REFUSED quiet-check: not quiet for 60 s straight within 600 s; now: ovv-audit-a-bat-p4-053313.scope (rule 4: a lane scope); ovv-p3-release-audit-B-bat | `S-HEAD/speed/QUIET-CHECK.jsonl` |
| S-HEAD | speed | a2 | quiet check (PASS) | 2026-10-03T05:57:04Z–05:59:25Z: quiet for 60 s: no busy unit or process, the 1-minute load average below 2 at all 13 polls (max 1.24) | `S-HEAD/speed/QUIET-CHECK.jsonl` |
| L-Q2 | speed | a2 | quiet check (REFUSED) | 2026-10-03T05:33:28Z–05:43:29Z: REFUSED quiet-check: not quiet for 60 s straight within 600 s; now: gao-audit-attack-1.scope (rule 4: a lane scope); ovv-audit-a-bat-p6-054233.scope ( | `L-Q2/speed/QUIET-CHECK.jsonl` |
| L-Q2 | speed | a2 | quiet check (PASS) | 2026-10-03T06:00:39Z–06:07:35Z: quiet for 60 s: no busy unit or process, the 1-minute load average below 2 at all 13 polls (max 1.29) | `L-Q2/speed/QUIET-CHECK.jsonl` |

### Restores (the parked timer read back after each attempt)

| arm | phase | attempt | who restored | timer read back | lock after | source |
|---|---|---:|---|---|---|---|
| L-IQ3 | quality | a1 | watcher | active | absent | `L-IQ3/quality/a1/RESTORE.json` |
| L-IQ3 | quality | a2 | watcher | active | absent | `L-IQ3/quality/a2/RESTORE.json` |
| S-IQ3 | speed | a1 | watcher | active | absent | `S-IQ3/speed/a1/RESTORE.json` |
| S-IQ3 | speed | a2 | watcher | active | absent | `S-IQ3/speed/a2/RESTORE.json` |
| S-IQ3 | quality | a1 | watcher | active | absent | `S-IQ3/quality/a1/RESTORE.json` |
| S-IQ3 | load | a1 | watcher | active | absent | `S-IQ3/load/a1/RESTORE.json` |
| S-IQ3-noMTP | load | a1 | watcher | active | absent | `S-IQ3-noMTP/load/a1/RESTORE.json` |
| S-Q2 | speed | a1 | watcher | active | absent | `S-Q2/speed/a1/RESTORE.json` |
| S-Q2 | speed | a2 | watcher | active | absent | `S-Q2/speed/a2/RESTORE.json` |
| S-Q2 | quality | a1 | watcher | active | absent | `S-Q2/quality/a1/RESTORE.json` |
| S-Q2 | load | a1 | watcher | active | absent | `S-Q2/load/a1/RESTORE.json` |
| L-Q2 | speed | a1 | watcher | active | absent | `L-Q2/speed/a1/RESTORE.json` |
| L-Q2 | speed | a2 | watcher | active | absent | `L-Q2/speed/a2/RESTORE.json` |
| L-Q2 | quality | a1 | watcher | active | absent | `L-Q2/quality/a1/RESTORE.json` |
| L-Q2 | fit | a1 | watcher | active | absent | `L-Q2/fit/a1/RESTORE.json` |
| L-Q2 | fit | a2 | watcher | active | absent | `L-Q2/fit/a2/RESTORE.json` |
| L-Q2 | fit | a3 | watcher | active | absent | `L-Q2/fit/a3/RESTORE.json` |
| L-Q2 | fit | a4 | watcher | active | absent | `L-Q2/fit/a4/RESTORE.json` |
| L-Q2 | fit | a5 | watcher | active | absent | `L-Q2/fit/a5/RESTORE.json` |
| S-HEAD | speed | a1 | watcher | active | absent | `S-HEAD/speed/a1/RESTORE.json` |
| S-HEAD | speed | a2 | watcher | active | absent | `S-HEAD/speed/a2/RESTORE.json` |
| S-HEAD | quality | a1 | watcher | active | absent | `S-HEAD/quality/a1/RESTORE.json` |
| S-HEAD | load | a1 | watcher | active | absent | `S-HEAD/load/a1/RESTORE.json` |
| L-IQ3-t8 | speed | a1 | watcher | active | absent | `L-IQ3-t8/speed/a1/RESTORE.json` |
| L-IQ3-fa0 (fa-off) | quality | a1 | watcher | active | absent | `L-IQ3-fa0/quality/a1/RESTORE.json` |
| L-IQ3-fa0 (fa-off) | load | a1 | watcher | active | absent | `L-IQ3-fa0/load/a1/RESTORE.json` |
| L-IQ3-fa0 (fa-off) | load | a2 | watcher | active | absent | `L-IQ3-fa0/load/a2/RESTORE.json` |
| L-Q2-fa0 (fa-off) | quality | a1 | watcher | active | absent | `L-Q2-fa0/quality/a1/RESTORE.json` |
| L-Q2-fa0 (fa-off) | load | a1 | watcher | active | absent | `L-Q2-fa0/load/a1/RESTORE.json` |
| L-Q2-fa0 (fa-off) | load | a2 | watcher | active | absent | `L-Q2-fa0/load/a2/RESTORE.json` |

## 2 · Speed (RESULT attempts only; client clock of record, PREREG A1-2)

A decode figure enters a median only when its token count holds (AUDIT-HARNESS B-1): COUNT-MISMATCH and COUNT-UNVERIFIED rows print `withheld` and their reason. The threads column is each start's own log line (PREREG A3).

| arm | ctx | threads | decode tok/s (median) | the repeats [row ids] | engine's own (median) | TTFT ms (median) | prompt tok/s, engine (median) | prompt tokens | completion tokens | count check (rows ok / recount gap max) | board J / token (median) | RAPL pkg J (not wall, median) | /data reads/s (mean of means) | VRAM growth MiB (max) | VmLck GiB (max) | hit rate (median) | tokens/round, derived (median) | flags |
|---|---|---|---:|---|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---|
| L-IQ3 | 4K | t 4 / tb 4 (server.log) | 20.2 | 19.3 [L-IQ3/speed/a1/run-001], 20.2 [L-IQ3/speed/a1/run-002], 20.5 [L-IQ3/speed/a1/run-003] | 20.2 | 15,199 | 265.8 | 4038 | 256 | 3/3 ok; recount gap max — | 6.755 | 1,106.3 | 66.5 | 22,452.0 | — | — | — | — |
| L-IQ3 | 32K | t 4 / tb 4 (server.log) | 18.4 | 17.7 [L-IQ3/speed/a1/run-004], 18.4 [L-IQ3/speed/a1/run-005], 18.5 [L-IQ3/speed/a1/run-006] | 18.4 | 126,715 | 252.9 | 32038 | 256 | 3/3 ok; recount gap max — | 37.231 | 5,125.8 | 827.2 | 22,454.0 | — | — | — | — |
| L-IQ3 | 128K | t 4 / tb 4 (server.log) | 13.7 | 13.6 [L-IQ3/speed/a1/run-007], 13.6 [L-IQ3/speed/a1/run-008], 13.9 [L-IQ3/speed/a1/run-009] | 13.7 | 557,645 | 229.6 | 128037 | 256 | 3/3 ok; recount gap max — | 150.909 | 24,009.0 | 1,153.7 | 22,454.0 | — | — | — | — |
| L-IQ3-t8 | 4K | t 8 / tb 8 (row) | 24.4 | 24.4 [L-IQ3-t8/speed/a1/run-001], 24.0 [L-IQ3-t8/speed/a1/run-002], 24.7 [L-IQ3-t8/speed/a1/run-003] | 24.4 | 19,083 | 211.7 | 4038 | 256 | 3/3 ok; recount gap max 0 | 6.931 | 1,333.8 | 1,687.1 | 22,452.0 | — | — | — | — |
| L-IQ3-t8 | 32K | t 8 / tb 8 (row) | 20.7 | 20.9 [L-IQ3-t8/speed/a1/run-004], 20.7 [L-IQ3-t8/speed/a1/run-005], 20.5 [L-IQ3-t8/speed/a1/run-006] | 20.7 | 125,598 | 255.2 | 32038 | 256 | 3/3 ok; recount gap max 1 | 36.818 | 5,113.5 | 14.7 | 22,454.0 | — | — | — | — |
| L-IQ3-t8 | 128K | t 8 / tb 8 (row) | 14.2 | 14.3 [L-IQ3-t8/speed/a1/run-007], 14.2 [L-IQ3-t8/speed/a1/run-008], 14.2 [L-IQ3-t8/speed/a1/run-009] | 14.2 | 545,113 | 234.9 | 128037 | 256 | 3/3 ok; recount gap max 0 | 149.439 | 19,817.8 | 845.3 | 22,454.0 | — | — | — | — |
| S-IQ3 · a1 · overlapped a lane's dry run; clean re-read pending (a2 STOPPED: memory) (window 2026-10-03T00:46:41Z–00:50:45Z; load1_mean 4.31–6.19 over 9 rows) | 4K | 23 workers + host (row) | 85.6 | 78.2 [S-IQ3/speed/a1/run-001], 91.4 [S-IQ3/speed/a1/run-002], 85.6 [S-IQ3/speed/a1/run-003] | 85.4 | 1,914 | 2,147.5 | 4038 | 256 | 3/3 ok; recount gap max 0 | 1.974 | 315.8 | 441.7 | 23,820.0 | — | 0.917 | 2.69 | — |
| S-IQ3 · a1 · overlapped a lane's dry run; clean re-read pending (a2 STOPPED: memory) (window 2026-10-03T00:46:41Z–00:50:45Z; load1_mean 4.31–6.19 over 9 rows) | 32K | 23 workers + host (row) | 87.0 | 87.3 [S-IQ3/speed/a1/run-004], 85.6 [S-IQ3/speed/a1/run-005], 87.1 [S-IQ3/speed/a1/run-006] | 86.9 | 12,131 | 2,655.1 | 32038 | 256 | 3/3 ok; recount gap max 1 | 7.674 | 639.3 | 141.0 | 23,820.0 | — | 0.932 | 2.56 | — |
| S-IQ3 · a1 · overlapped a lane's dry run; clean re-read pending (a2 STOPPED: memory) (window 2026-10-03T00:46:41Z–00:50:45Z; load1_mean 4.31–6.19 over 9 rows) | 128K | 23 workers + host (row) | 69.8 | 71.9 [S-IQ3/speed/a1/run-007], 69.8 [S-IQ3/speed/a1/run-008], 64.9 [S-IQ3/speed/a1/run-009] | 69.7 | 47,557 | 2,705.0 | 128037 | 256 | 3/3 ok; recount gap max 0 | 28.423 | 2,551.5 | 4,942.0 | 23,820.0 | — | 0.923 | 2.25 | — |
| S-Q2 · a1 · overlapped a lane's dry run, not the figure of record (window 2026-10-03T00:40:23Z–00:44:02Z; load1_mean 2.03–5.97 over 9 rows) | 4K | 23 workers + host (row) | 126.1 | 116.2 [S-Q2/speed/a1/run-001], 131.5 [S-Q2/speed/a1/run-002], 126.1 [S-Q2/speed/a1/run-003] | 125.8 | 1,746 | 2,354.1 | 4038 | 256 | 3/3 ok; recount gap max 0 | 1.420 | 176.6 | 459.9 | 23,796.0 | — | 0.983 | 2.61 | — |
| S-Q2 · a1 · overlapped a lane's dry run, not the figure of record (window 2026-10-03T00:40:23Z–00:44:02Z; load1_mean 2.03–5.97 over 9 rows) | 32K | 23 workers + host (row) | 120.8 | 120.8 [S-Q2/speed/a1/run-004], 120.3 [S-Q2/speed/a1/run-005], 127.0 [S-Q2/speed/a1/run-006] | 120.6 | 11,437 | 2,818.4 | 32038 | 256 | 3/3 ok; recount gap max 0 | 6.677 | 489.8 | 170.3 | 23,796.0 | — | 0.985 | 2.56 | — |
| S-Q2 · a1 · overlapped a lane's dry run, not the figure of record (window 2026-10-03T00:40:23Z–00:44:02Z; load1_mean 2.03–5.97 over 9 rows) | 128K | 23 workers + host (row) | 110.5 | 106.6 [S-Q2/speed/a1/run-007], 112.8 [S-Q2/speed/a1/run-008], 110.6 [S-Q2/speed/a1/run-009] | 110.3 | 45,166 | 2,848.5 | 128037 | 256 | 3/3 ok; recount gap max 0 | 26.732 | 1,871.3 | 5,445.7 | 23,796.0 | — | 0.991 | 2.42 | — |
| S-Q2 · a2 · clean re-read (PREREG A5-3), the figure of record (window 2026-10-03T05:09:47Z–05:13:25Z; load1_mean 1.83–7.07 over 9 rows; quiet check PASS 2026-10-03T05:09:46Z) | 4K | 23 workers + host (row) | 131.7 | 120.3 [S-Q2/speed/a2/run-001], 134.9 [S-Q2/speed/a2/run-002], 131.7 [S-Q2/speed/a2/run-003] | 131.3 | 1,717 | 2,399.3 | 4038 | 256 | 3/3 ok; recount gap max 0 | 1.353 | 213.0 | 447.9 | 23,796.0 | — | 0.983 | 2.64 | OUT-OF-BAND (opened: no) |
| S-Q2 · a2 · clean re-read (PREREG A5-3), the figure of record (window 2026-10-03T05:09:47Z–05:13:25Z; load1_mean 1.83–7.07 over 9 rows; quiet check PASS 2026-10-03T05:09:46Z) | 32K | 23 workers + host (row) | 115.2 | 114.7 [S-Q2/speed/a2/run-004], 115.2 [S-Q2/speed/a2/run-005], 120.9 [S-Q2/speed/a2/run-006] | 115.1 | 11,454 | 2,814.1 | 32038 | 256 | 3/3 ok; recount gap max 0 | 7.135 | 808.2 | 232.7 | 23,796.0 | — | 0.988 | 2.48 | — |
| S-Q2 · a2 · clean re-read (PREREG A5-3), the figure of record (window 2026-10-03T05:09:47Z–05:13:25Z; load1_mean 1.83–7.07 over 9 rows; quiet check PASS 2026-10-03T05:09:46Z) | 128K | 23 workers + host (row) | 102.5 | 106.4 [S-Q2/speed/a2/run-007], 102.5 [S-Q2/speed/a2/run-008], 102.2 [S-Q2/speed/a2/run-009] | 102.3 | 45,239 | 2,843.2 | 128037 | 256 | 3/3 ok; recount gap max 0 | 26.297 | 2,520.8 | 5,355.8 | 23,796.0 | — | 0.989 | 2.27 | — |
| L-Q2 · a1 · overlapped a lane's dry run, not the figure of record (window 2026-10-03T00:16:35Z–00:39:09Z; load1_mean 0.99–2.51 over 9 rows) | 4K | t 8 / tb 8 (row) | 30.6 | 30.6 [L-Q2/speed/a1/run-001], 29.8 [L-Q2/speed/a1/run-002], 31.5 [L-Q2/speed/a1/run-003] | 30.6 | 9,225 | 438.0 | 4038 | 256 | 3/3 ok; recount gap max 0 | 5.074 | 725.5 | 25.0 | 22,826.0 | — | — | — | — |
| L-Q2 · a1 · overlapped a lane's dry run, not the figure of record (window 2026-10-03T00:16:35Z–00:39:09Z; load1_mean 0.99–2.51 over 9 rows) | 32K | t 8 / tb 8 (row) | 25.3 | 24.8 [L-Q2/speed/a1/run-004], 26.3 [L-Q2/speed/a1/run-005], 25.3 [L-Q2/speed/a1/run-006] | 25.3 | 76,652 | 418.1 | 32038 | 256 | 3/3 ok; recount gap max 0 | 27.889 | 3,440.8 | 20.9 | 22,828.0 | — | — | — | — |
| L-Q2 · a1 · overlapped a lane's dry run, not the figure of record (window 2026-10-03T00:16:35Z–00:39:09Z; load1_mean 0.99–2.51 over 9 rows) | 128K | t 8 / tb 8 (row) | 17.2 | 16.9 [L-Q2/speed/a1/run-007], 17.2 [L-Q2/speed/a1/run-008], 18.1 [L-Q2/speed/a1/run-009] | 17.2 | 316,130 | 405.1 | 128037 | 256 | 3/3 ok; recount gap max 0 | 111.305 | 11,857.0 | 52.3 | 22,828.0 | — | — | — | — |
| L-Q2 · a2 · clean re-read (PREREG A5-3), the figure of record (window 2026-10-03T06:07:35Z–06:31:09Z; load1_mean 1.25–2.46 over 9 rows; quiet check PASS 2026-10-03T06:07:35Z) | 4K | t 8 / tb 8 (row) | 32.5 | 30.0 [L-Q2/speed/a2/run-001], 32.5 [L-Q2/speed/a2/run-002], 33.4 [L-Q2/speed/a2/run-003] | 32.5 | 9,487 | 426.0 | 4038 | 256 | 3/3 ok; recount gap max 0 | 5.079 | 696.9 | 758.1 | 22,824.0 | — | — | — | — |
| L-Q2 · a2 · clean re-read (PREREG A5-3), the figure of record (window 2026-10-03T06:07:35Z–06:31:09Z; load1_mean 1.25–2.46 over 9 rows; quiet check PASS 2026-10-03T06:07:35Z) | 32K | t 8 / tb 8 (row) | 26.6 | 26.6 [L-Q2/speed/a2/run-004], 27.6 [L-Q2/speed/a2/run-005], 26.1 [L-Q2/speed/a2/run-006] | 26.6 | 73,314 | 437.2 | 32038 | 256 | 3/3 ok; recount gap max 0 | 27.221 | 2,738.9 | 55.4 | 22,826.0 | — | — | — | — |
| L-Q2 · a2 · clean re-read (PREREG A5-3), the figure of record (window 2026-10-03T06:07:35Z–06:31:09Z; load1_mean 1.25–2.46 over 9 rows; quiet check PASS 2026-10-03T06:07:35Z) | 128K | t 8 / tb 8 (row) | 17.5 | 17.9 [L-Q2/speed/a2/run-007], 17.5 [L-Q2/speed/a2/run-008], 17.0 [L-Q2/speed/a2/run-009] | 17.5 | 321,839 | 398.0 | 128037 | 256 | 3/3 ok; recount gap max 0 | 110.837 | 15,658.1 | 5.2 | 22,826.0 | — | — | — | — |
| S-HEAD · a1 · overlapped a lane's dry run; clean re-read pending (a2 STOPPED: memory) (window 2026-10-03T00:55:33Z–00:59:30Z; load1_mean 3.97–8.43 over 9 rows) | 4K | 23 workers + host (row) | 80.4 | 80.0 [S-HEAD/speed/a1/run-001], 80.4 [S-HEAD/speed/a1/run-002], 83.6 [S-HEAD/speed/a1/run-003] | 80.4 | 1,793 | 2,299.4 | 4038 | 256 | 3/3 ok; recount gap max 0 | 1.952 | 289.0 | 435.1 | 23,732.0 | — | 0.915 | 2.67 | — |
| S-HEAD · a1 · overlapped a lane's dry run; clean re-read pending (a2 STOPPED: memory) (window 2026-10-03T00:55:33Z–00:59:30Z; load1_mean 3.97–8.43 over 9 rows) | 32K | 23 workers + host (row) | 79.6 | 79.6 [S-HEAD/speed/a1/run-004], 78.6 [S-HEAD/speed/a1/run-005], 92.1 [S-HEAD/speed/a1/run-006] | 79.4 | 11,250 | 2,865.9 | 32038 | 256 | 3/3 ok; recount gap max 0 | 7.214 | 759.4 | 125.3 | 23,732.0 | — | 0.933 | 2.54 | — |
| S-HEAD · a1 · overlapped a lane's dry run; clean re-read pending (a2 STOPPED: memory) (window 2026-10-03T00:55:33Z–00:59:30Z; load1_mean 3.97–8.43 over 9 rows) | 128K | 23 workers + host (row) | 74.0 | 74.0 [S-HEAD/speed/a1/run-007], 72.3 [S-HEAD/speed/a1/run-008], 75.3 [S-HEAD/speed/a1/run-009] | 73.9 | 46,114 | 2,789.5 | 128037 | 256 | 3/3 ok; recount gap max 0 | 27.574 | 2,373.3 | 5,163.7 | 23,732.0 | — | 0.921 | 2.35 | — |

### The registered bands at 4K (PREREG section 2)

- L-IQ3 4K: 20.2 tok/s inside its band 20–55 — rows L-IQ3/speed/a1/run-001, L-IQ3/speed/a1/run-002, L-IQ3/speed/a1/run-003
- L-IQ3-t8 4K: 24.4 tok/s inside its band 20–55 — rows L-IQ3-t8/speed/a1/run-001, L-IQ3-t8/speed/a1/run-002, L-IQ3-t8/speed/a1/run-003
- S-IQ3 4K, a1 (overlapped a lane's dry run; clean re-read pending (PREREG A5-3)): 85.6 tok/s inside its band 40–90 — rows S-IQ3/speed/a1/run-001, S-IQ3/speed/a1/run-002, S-IQ3/speed/a1/run-003
- **S-Q2 4K, a2 (the clean re-read, the figure of record (PREREG A5-3)): 131.7 tok/s is OUTSIDE its band 60–130: OUT-OF-BAND (opened: no)** — rows S-Q2/speed/a2/run-001, S-Q2/speed/a2/run-002, S-Q2/speed/a2/run-003
  - beside it a1 (overlapped a lane's dry run; load1_mean 2.03–2.70 over these rows): 126.1 tok/s — rows S-Q2/speed/a1/run-001, S-Q2/speed/a1/run-002, S-Q2/speed/a1/run-003
- L-Q2 4K, a2 (the clean re-read, the figure of record (PREREG A5-3)): 32.5 tok/s inside its band 10–45 — rows L-Q2/speed/a2/run-001, L-Q2/speed/a2/run-002, L-Q2/speed/a2/run-003
  - beside it a1 (overlapped a lane's dry run; load1_mean 0.99–2.11 over these rows): 30.6 tok/s — rows L-Q2/speed/a1/run-001, L-Q2/speed/a1/run-002, L-Q2/speed/a1/run-003
- S-HEAD 4K, a1 (overlapped a lane's dry run; clean re-read pending (PREREG A5-3)): 80.4 tok/s — no registered band (unregistered arm) — rows S-HEAD/speed/a1/run-001, S-HEAD/speed/a1/run-002, S-HEAD/speed/a1/run-003

### PREREG A7 · the llama.cpp controls beside L-IQ3-t8

L-IQ3-ub, L-IQ3-t16: controls with no band and no gate; nothing here feeds G0, G1, G2 or "Better". Each condition is read from that attempt's own argv, each n_batch / n_ubatch from that start's own server.log (src/llama-context.cpp:309-310, after the clamps), and each N from the argv with the frozen-N file the marker names. The first-token ratio is the base's median over this arm's median, from the medians printed.

| ctx | arm | condition (argv) | n_batch / n_ubatch (server.log) | --n-cpu-moe (source) | threads | TTFT ms (median) | first-token ratio to the base | prompt tok/s, engine (median) | decode tok/s (median) | the repeats' decode [row ids] |
|---|---|---|---|---|---|---:|---|---:|---:|---|
| 4K | L-IQ3-t8 · a1 | no -b / -ub (llama.cpp's own 2048 / 512); -t 8 -tb 8 | n_batch = 2048, n_ubatch = 512 | 34 (FROZEN-NCMOE-L-IQ3.json) | t 8 / tb 8 (row) | 19,083 | the base | 211.7 | 24.4 | 24.4 [L-IQ3-t8/speed/a1/run-001], 24.0 [L-IQ3-t8/speed/a1/run-002], 24.7 [L-IQ3-t8/speed/a1/run-003] |
| 32K | L-IQ3-t8 · a1 | no -b / -ub (llama.cpp's own 2048 / 512); -t 8 -tb 8 | n_batch = 2048, n_ubatch = 512 | 34 (FROZEN-NCMOE-L-IQ3.json) | t 8 / tb 8 (row) | 125,598 | the base | 255.2 | 20.7 | 20.9 [L-IQ3-t8/speed/a1/run-004], 20.7 [L-IQ3-t8/speed/a1/run-005], 20.5 [L-IQ3-t8/speed/a1/run-006] |
| 128K | L-IQ3-t8 · a1 | no -b / -ub (llama.cpp's own 2048 / 512); -t 8 -tb 8 | n_batch = 2048, n_ubatch = 512 | 34 (FROZEN-NCMOE-L-IQ3.json) | t 8 / tb 8 (row) | 545,113 | the base | 234.9 | 14.2 | 14.3 [L-IQ3-t8/speed/a1/run-007], 14.2 [L-IQ3-t8/speed/a1/run-008], 14.2 [L-IQ3-t8/speed/a1/run-009] |

- L-IQ3-ub speed: no attempt yet
  - frozen N: not frozen yet
- L-IQ3-t16 speed: no attempt yet (optional, PREREG A7-6)
  - N: read from FROZEN-NCMOE-L-IQ3.json (no fit of its own)

## 3 · Quality (parser v2 strict; G1, G2, Better)

Reference check (re-derived from the laptop rows at estate `7d2a37bc`): qwen3.8:27b six-way 93/108, the gemma4 anchor's doorman 36/48, qwen3.5:27b 96/108 and 46/48.

**PREREG A4 (fa-off).** G1's llama.cpp side is read from the fa-off rows and printed with that condition: L-IQ3-fa0 (fa-off) stands for L-IQ3 against S-IQ3; L-Q2-fa0 (fa-off) stands for L-Q2 against S-Q2. If an fa-off arm also fails inside the doorman task, its doorman reads `not read (engine-failed)` and so does G1's doorman half; its six-way rows still read. The fa-on quality attempts stay as recorded: L-IQ3 a1 STOPPED, a2 STOPPED; L-Q2 a1 STOPPED.

**PREREG A5 (the fa-off condition).** Every fa-off row below, and every G1 cell read against one, prints its condition from that attempt's own argv: `fa-off (V cache f16, forced) at an 8K context` on A5-2's line (`-c 8192`; the V cache is f16 because 3cf03257 refuses a quantized V cache without flash attention, which also turns off V's rotation, A5-1), and the context the argv names on any other line. The two loads at A4's line (`-c 131072`) stopped before /health on the compute buffer's cudaMalloc (A5-2).

| arm | six-way strict | doorman strict | G2 (>= 93 and >= 36) | G1 vs its pair (same strict choice; >= 105 and >= 47) | Better vs qwen3.5:27b (exact McNemar) | source rows |
|---|---:|---:|---|---|---|---|
| L-IQ3-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | 91/108 (108 rows) | 7/48 (48 rows) | fail | miss (disclosed; the two engines' rows stand separately): c 108/108 (>= 105: yes), d 44/48 (>= 47: no) vs S-IQ3 | no win claimed (b=3, c=47, p=0.0) | L-IQ3-fa0/quality/a1/task-c-001, L-IQ3-fa0/quality/a1/task-d-001 |
| S-IQ3 | 91/108 (108 rows) | 11/48 (48 rows) | fail | miss (disclosed; the two engines' rows stand separately): c 108/108 (>= 105: yes), d 44/48 (>= 47: no) vs L-IQ3-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | no win claimed (b=3, c=43, p=0.0) | S-IQ3/quality/a1/task-c-001, S-IQ3/quality/a1/task-d-001 |
| S-Q2 | 88/108 (108 rows) | 0/48 (48 rows) | fail | miss (disclosed; the two engines' rows stand separately): c 103/108 (>= 105: no), d 48/48 (>= 47: yes) vs L-Q2-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | no win claimed (b=5, c=59, p=0.0) | S-Q2/quality/a1/task-c-001, S-Q2/quality/a1/task-d-001 |
| L-Q2-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | 86/108 (108 rows) | 0/48 (48 rows) | fail | miss (disclosed; the two engines' rows stand separately): c 103/108 (>= 105: no), d 48/48 (>= 47: yes) vs S-Q2 | no win claimed (b=5, c=61, p=0.0) | L-Q2-fa0/quality/a1/task-c-001, L-Q2-fa0/quality/a1/task-d-001 |
| S-HEAD | 91/108 (108 rows) | 8/48 (48 rows) | fail | miss (disclosed; the two engines' rows stand separately): c 108/108 (>= 105: yes), d 41/48 (>= 47: no) vs L-IQ3-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | no win claimed (b=3, c=46, p=0.0) | S-HEAD/quality/a1/task-c-001, S-HEAD/quality/a1/task-d-001 |

A rung that fails G2 keeps its speed rows but gets no value table (PREREG section 2). "Better" is the only win anyone may claim.

### Attention per llama.cpp quality and load attempt (each start's own server.log; PREREG A4)

| arm | phase | attempt | marker | flash_attn (log) | K / V cache (log) | -fa / -ctv (argv) | source |
|---|---|---:|---|---|---|---|---|
| L-IQ3 | quality | a1 | STOPPED | enabled | q8_0 / q8_0 | on / q8_0 | `L-IQ3/quality/a1/server.log` |
| L-IQ3 | quality | a2 | STOPPED | enabled | q8_0 / q8_0 | on / q8_0 | `L-IQ3/quality/a2/server.log` |
| L-IQ3-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | quality | a1 | RESULT | disabled | q8_0 / f16 | off / f16 | `L-IQ3-fa0/quality/a1/server.log` |
| L-IQ3-fa0 (fa-off) · fa-off (V cache f16, forced) at a 128K context | load | a1 | STOPPED | disabled | q8_0 / f16 | off / f16 | `L-IQ3-fa0/load/a1/server.log` |
| L-IQ3-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | load | a2 | RESULT | disabled | q8_0 / f16 | off / f16 | `L-IQ3-fa0/load/a2/server.log` |
| L-Q2 | quality | a1 | STOPPED | enabled | q8_0 / q8_0 | on / q8_0 | `L-Q2/quality/a1/server.log` |
| L-Q2-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | quality | a1 | RESULT | disabled | q8_0 / f16 | off / f16 | `L-Q2-fa0/quality/a1/server.log` |
| L-Q2-fa0 (fa-off) · fa-off (V cache f16, forced) at a 128K context | load | a1 | STOPPED | disabled | q8_0 / f16 | off / f16 | `L-Q2-fa0/load/a1/server.log` |
| L-Q2-fa0 (fa-off) · fa-off (V cache f16, forced) at an 8K context | load | a2 | RESULT | disabled | q8_0 / f16 | off / f16 | `L-Q2-fa0/load/a2/server.log` |

