The weekend's two runs: CoactDetect holds under the budget, and the rest is inside the noise

[syncytium2]

Rewritten against the final report, which #665 landed while this PR was open. Three blind murderboard rounds moved the headline twice, and the first version of this page carried the round-1 reading. Corrected here before it reaches main.

What holds

Under the shared false-alarm budget — 1.6× CoactDetect's own rates — CoactDetect is ahead of every net in every fold of both draws. That is the selection with a stated operating constraint, and it is the one that replicated.

What does not

Chosen on F1 alone the two are nearly tied and the sign is not settled. Counting every training, CoactDetect leads on average in both draws. But the best net moves ahead once both sides use the same merge gap — a setting only the coded side was allowed to tune — and in the replicate it leads in most folds, its average pulled down by two failed trainings. The matched-gap leads carry t of only −1.5 to −2.1, and margins this small are at the limit of what the scoring resolves.

The replicate supplied the scale that makes any of this checkable: changing only the recordings moves a net by a median of 0.010 F1, a coded detector by 0.003. The shakedown's leads of +0.011 and +0.016 were exactly that size — a single run could never have separated them from the draw.

Three findings a single pass would have missed

1. The earlier +0.103 lead is gone even for the untuned nets, so it was not lost to tuning them. The simulator and CoactDetect's own tuning changed together and this run cannot separate them.
2. The two sides win in different places. On F1 alone the chorus nets take the faintest events — a tenth of the cells — when background firing is high; CoactDetect takes them when it is low. A pooled F1 hides that entirely.
3. The nets are handicapped by construction: each fits 10 of 72 training recordings and picks its threshold on 2, while a coded setting is scored on all 72.

Corrections to the earlier version of this PR

• Binned SCE does beat CoactDetect under the budget in a few folds (3 of 4 replicate folds). Only its F1-alone first place fails — a 30 s merge gap it moved to in 8 of 8 folds, failing the crowded check in 7 of 8. The earlier text said its win did not stand, full stop.
• The matched-gap reversal was stated as though decisive. It applies to F1 alone, is weak, and under the budget nothing reverses.
• 064 ran three murderboard rounds and a follow-up pass, not one.

Still recorded: both workstations re-measured the bench on the de-pinned export on 2026-09-17, every value inside its bootstrap interval (2120516, not on main).

chorus_norm's training collapse is under diagnosis on WSMIP065; the merge-gap rerun is with WSMIP064.

Docs only, two files. tests/test_index_resolves.py 329 passed; sapper, check_quotes and check_pipelines clear.

🤖 Generated with Claude Code

https://claude.ai/code/session_017hQX6iESxBsQk7Jan875e3