Lab / SEP 17 2026 / note
Two arms drifted +11 ms together while every gate read ok
A previous campaign's arms co-drifted by 11 ms with every environment gate reporting OK. The response was not a better gate but an interleaved protocol with the per-round series printed.
Measurement methodology, Contaminated runs, ControlsSAM3- Machine
- RTX 5090, Vast.ai instance 51353366
- Commit
- `gabilan/sam3-ggml` `ewi1963/f16-vectorized-converts@429236ef` (kernels); the harness has no commit
- Co-drift in the previous campaign
- +11 ms, both arms together, every gate reading OK
- Interference variation in this campaign
- other_cpus 3.59 → 10.20
The header of `run_ab.sh` says why the harness looks the way it does. Two arms of a previous campaign drifted +11 ms *together* while every gate read ok.
The quiet gate gates CPU, not the GPU: two arms of a previous campaign drifted +11 ms TOGETHER while every gate read ok (ewi1963-2x-push s4.1). So this is interleaved and balanced (A B | B A | ...) with one log per round, and it prints the per-round series so drift is visible in the data rather than inferred from a gate reading.
Run those two arms without interleaving and the co-drift is a published regression. Both arms would have moved, the difference between them would have been unchanged, and the numbers on the page would have said the machine got slower. A gate would not have caught it, because the gate was fine.
The response was not a better gate
The fix was a protocol that does not depend on the gate: interleave the arms, rotate their position inside each round, give each arm its own source tree and binary, and print the per-round series so drift shows up in the numbers. The environment gate stays, as a necessary and insufficient instrument.
This is Weft's benchmark rule 7 reached independently — that no gate on the host can detect a contaminated run. Two projects, two machines, the same conclusion.