Date: 2026-07-23 America/Toronto
The down-only native-M8 shared-expert treatment passed the frozen four-card component gate. This is a component result, not an endpoint benchmark or a LocalMaxxing record.
The passing root is:
/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/shared-down-m8-component-20260723T155703Z
The aggregate artifact is aggregate.json, SHA-256
ea71971b368ce9b9e930577b673e983124b0e5686d5d780fc241ac4104f2a1d6.
The frozen analyzer classified it
component-passed-counter-tooling-freeze-next.
| Rank | BDF | Wins | Control ms/cycle | Candidate ms/cycle | Saving ms/cycle | Relative |
|---|---|---|---|---|---|---|
| 0 | 0000:23:00.0 |
31/31 | 2.309867 | 1.698725 | 0.610891 | 26.4470% |
| 1 | 0000:27:00.0 |
31/31 | 2.391377 | 1.745049 | 0.647105 | 27.0599% |
| 2 | 0000:43:00.0 |
31/31 | 2.313656 | 1.715529 | 0.597514 | 25.8255% |
| 3 | 0000:47:00.0 |
31/31 | 2.361052 | 1.771314 | 0.612379 | 25.9367% |
Every card independently exceeded the frozen requirements of at least 28/31 block wins and at least 0.15 ms median saving per complete 47-layer cycle. The narrowest observed margin was rank 2 at 0.597514 ms, nearly four times the absolute threshold.
Each card passed 128 changing exactness epochs, the actual checkpoint-selected
RowParallelLinear path, and a 32-epoch post-timing replay. The candidate,
candidate repeat, shared+routed add, and fixed-rank simulated sum were raw
bitwise equal to control at every declared boundary. The marked M=8 path
issued exactly one native MM; unmarked M=8 and marked M=7 issued exactly two
incumbent BMMs; the bad-layout candidate failed closed.
All four cards reported the same aggregate fixture hash
3e28840809747843474a15f7858db9b7d1d4d70b4fbe71c47c7a2aa117eeff90
and output hash
ae8c34ea1bb5904466a702412a1ccc1f6843d3bed05e948f079e82647b4f33a7.
The analyzer verified four distinct UUIDs/BDFs, one boot, one clean main
commit, and the frozen vLLM, kernel, binary, configuration, environment, and
tool identities.
The captured source identities were:
480406187af3ae01ca1f4eaedf4810fee6c1ecfd;75d4660463407975c16bd33711499ca560bf2034;c59aaadbbfd350c2b5f4ad663e247c2811ae3181;df8496f1f405e8b786dff0b96b7c320944c5d0133cce0bfcc2e36150ab1e0f12;
and945810c50eeeea99f532c3e62ee5bf289677e3706d80965f966400bfab35911b.The structured summary is
data/laguna-s-2.1-shared-down-m8-component-pass-20260723.json.
The full protocol and the two sealed preflight-tool failures remain in the
preregistration.
This pass authorizes only construction, source freezing, validation, and independent audit of dedicated cold hardware-counter tooling. It does not authorize counter execution, an endpoint service, model generation, a payload, or a LocalMaxxing submission. Those flags remain false in every card artifact and the aggregate.
The next step is to implement the preregistered two-pair per-card cold
unitrace gate, freeze its source and hashes, and audit it before running any
counter capture. All live evidence remains on local NVMe; the external USB
remains backup-only.