Date: 2026-07-24 America/Toronto
Status: the first and only authorized cold-counter capture completed all sixteen arms and is terminal before endpoint work. The candidate remained bitwise exact, but the frozen parser failed and the immutable counter bytes also fail the core matched-pair and per-card promotion requirements. There is no rerun, acceptance repair, endpoint, model generation, payload, network access, record claim, reboot, or LocalMaxxing submission.
The counter tooling is commit
34db11e8f9cee45e455390da7961e28c959b0441. The immediate packet-only
authorization is commit
a8c8c595978e1803a354869d53cef77cae79781c; its packet SHA-256 is
a91749870de477207d21c464660adde164cf1c9b24faca2e62d836a4046b4748.
The complete live root is on internal NVMe:
/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/shared-gate-up-m8-counters-20260724T063300Z
The one-shot runner completed physical cards 0 through 3 sequentially, with
A1 -> B1 -> B2 -> A2 on every card. Each arm contains exactly 13
completion-bounded gate-then-up pairs, exact PID-bound unitrace timing and
ComputeBasic files, and a closed manifest. The principal seals are:
591aeb60b3fddf42006f1152a7ee307c6f814145f64c1cb9a10c68de29e8a7c1;473ec7f973a0f3ecf86f7cf0c17dcaf38ff4828c1ed4622988f70f4523393516;0c082d4c3ce71987587ba5c8293ad6348d750840a3084121714a581baa943654;
and7f6738d0b8c75b43e63ced5208a89cadfe7f9c6fff63ec423c38133fc8ed9b84.The external USB was not used.
The fixture-level exactness closure passed. Every one of the sixteen arms repeated the same raw BF16 gate and up hashes on all 13 pairs, and the cross-arm/card tuple count is one:
49fd9c1dc22472e942bfae8031b8b8e5aa08ffb85b98158edea6c3134d16bd37;bf117165cb35ce114581e127522ba287b9440b7ef88ea8ed068647556edc5e40;
and0f005ce34fbf306b9177b2f70043cb0ef4ad8b3472d3a40774f25676f4b0be32.The sealed component predecessor still covers the nine downstream BF16
boundaries with final-manifest SHA-256
8aa2a45a8bcc31c5d2e84e5f55568ad43144ebd6c380d6d914cf91f77d10a10d.
Quality was not sacrificed.
The offline analyzer stopped at the first card-0 A1 metric row:
nonzero invalidity/traffic proxy SLM_BYTE_READ[bytes]
All 416 selected metric rows actually report exactly 245,760 SLM bytes read and 245,760 SLM bytes written, for both control and candidate. Every row has zero SLM bank conflicts, zero LSC partial writes, and zero LSC byte writes. This shows the preregistered parser misclassified ordinary constant SLM traffic as a zero-required failure proxy.
That observation does not authorize changing the acceptance rule after
seeing the result. The parser and packet were frozen before capture, so the
campaign remains failed and analysis.error.json remains immutable. No
offline repair is used to manufacture a pass.
For research direction only, the retained metric rows were summarized without treating that summary as a replacement analyzer decision. Even if the SLM rule were ignored, the candidate fails the original no-rescue requirements:
| Rank | B1/A1 GPU time | B2/A2 GPU time | Candidate/control aggregate | Occupancy delta pp | XVE-stall delta pp |
|---|---|---|---|---|---|
| 0 | 0.994767 | 0.633845 | 0.812314 | +0.798427 | +1.105145 |
| 1 | 0.937180 | 1.401331 | 1.163663 | +1.105352 | +0.645766 |
| 2 | 1.056653 | 0.815515 | 0.921325 | +1.741008 | +1.650995 |
| 3 | 0.956882 | 1.320946 | 1.117919 | -2.450955 | -2.292902 |
Required matched comparisons fail at rank 1 B2/A2, rank 2 B1/A1, and rank 3
B2/A2. Card aggregates fail on ranks 1 and 3. Complete per-card guardrails
also fail: ranks 0 through 2 exceed the 0.5 percentage-point XVE-stall
allowance, and rank 3 loses 2.451 occupancy percentage points. The global
candidate/control GPU-time ratio is 0.9935168, a diagnostic 0.6483% win,
but global evidence is explicitly forbidden from rescuing any failed pair,
card, or guardrail.
The component loop’s 9.4-11.2% repeated-cycle gain therefore does not survive the cold per-dispatch variance contract. This is useful negative evidence, not a promotion result.
The separate shared gate+up native-M8 promotion lane is closed. The capture must not be rerun, selectively sampled, reinterpreted, or used to authorize an endpoint. Preserve the candidate and this evidence as a measured negative. The next optimization must be materially different and preregistered before its evidence is observed.
The structured summary is
data/laguna-s-2.1-shared-gate-up-m8-counter-terminal-negative-20260724.json.