The nets' merge gap, tuned like any other setting: the answer holds, and the check bounds both sides

[syncytium2]

The merge gap the nets were never allowed to tune, tuned by the same rules — on both draws, with no retraining — and murderboarded twice.

What it settles. In the fair comparison the coded detectors' searches tuned their merge gap (CoactDetect took 8 s in every fold) while the nets decoded at the fixed 2 s their threshold picker returns. That was the open asymmetry. Here the nets' gap is selected, on the inner fits of the training folds, by the rules the run used for everything else, with goal 1's move rule and its crowded-recording check inside the search.

No retraining, and it reproduces the run exactly. A merge gap is applied when a net's per-frame output is decoded, so every saved fit of both draws — 3,881 of them — was re-run and decoded at each gap in the grid. At 2 s every picked threshold, every recording's counts, every empty recording's call count and both selections' inner and held-out F1 match the runs' own files.

The result

• Under the budget the answer holds. CoactDetect is ahead of every net on average in both draws and in 29 of 32 net-folds. Tuning closes 0.009 to 0.014 F1 of the chorus nets' shortfall.
• On F1 alone the sign depends on the accounting, and the gap is worth about one noise unit either way. Scored the same way on both sides, the replicate's chorus nets were already ahead at 2 s and tuning adds ~0.010 F1.
• Every chosen gap is a boundary, not an optimum. Held-out F1 peaks at the widest gap in all 64 fold-rows; the nets stop where the crowded check stops them, and CoactDetect's 8 s is the top of its own grid — which stops there because goal 1's check refused its 16 s. Both sides are bounded by the same check, at different times.
• The long-gap inflation is not specific to the coded side. In all 64 choices the training folds preferred a wider gap and the check refused it, 30 s for every net.
• ⚠ The accepted settings are not free: a median 0.012 F1 lost on crowded recordings, the same size as the gain, with two choices failing the check outright on the outer refits. On those same recordings, head to head, the tuned chorus nets beat CoactDetect while tube loses 0.14 — the comparison the cost is not, now drawn.

The review, and why this ships unconverged. Eleven roles blind, then five re-run blind against the rebuilt page: 11 blocking findings in round 1, 10 in round 2. The process says a flat blocking count means a structural problem patching will not retire, and to escalate rather than run a third round — so both rounds' fixes are applied and the record ships with named residuals rather than a clean bill. The review changed what the page may claim, not just how it reads: the one-sided set-aside, the censored boundary, the mis-referenced crowded cost, the noise scale's own systematic shift, and a budget mechanism the data refute.

What is here

• tools/tune_net_merge_gap.py: the re-decoding and the selections, reusing the tuning tool's own objective, budget, pooling and tie rules. Records the machine, torch and both trees' commits; asserts the threshold grid it indexes.
• tools/build_net_merge_gap_page.py: the page. Every sentence about the data is computed, and the directional ones sit behind claim(), which stopped the build three times during this work.
• docs/reviews/net-merge-gap-2026-09-19.md plus both role archives, verbatim.
• docs/goals/learned-model-family.md, docs/GLOSSARY.md: updated in the same PR, as the rules require.
• docs/sapper_feedback/2026-09-19-the-roster-gate-eats-underscores.md: the roster gate rewrites underscores in the archive path and reports a complete archive as missing.

Open for Tony: the 0.02 F1 crowded allowance is still unsigned and decides every gap here — one refusal turned on 0.0009 F1.

🤖 Generated with Claude Code