Two draws of goal 2's comparison: the report, and what the pair can claim

[syncytium2]

WSMIP065 reran goal 2's learned-versus-coded comparison on recording seeds 2000–2047, disjoint from WSMIP064's first draw (1000–1047), with everything else held fixed. This PR adds the report for a reader new to the project, the tool that builds it, and the review record.

What the pair of draws says (every number is asserted against the run files at build time)

• Changing only the recordings moves a score by about a hundredth of F1: a median of 0.010 for the nets and 0.003 for the coded detectors.
• Under the shared false-alarm budget, every learned model trails CoactDetect in every fold of both draws. The closest, chorus_gain_norm, trails on average by two to three times the typical move (paired t: p = 0.001 and 0.07).
• Part of that gap is the merge setting. The nets join calls within 2 s and CoactDetect within 8 s. Re-decoded with both sides at the same gap, using WSMIP064's fair_comparison_evidence.py merge-gap run unchanged on this draw, 43–72% of the gap remains. On F1 alone the chorus models come out at or slightly above CoactDetect, but that is not established.
• All of the recall difference is in the 3-cell events.
• Goal 1's crowded-recording check, run unchanged on both draws, refuses these coded choices in every fold: binned SCE chosen on F1 alone, and LoCo, rate+context and SPIKE-synch chosen under the budget. CoactDetect passes everywhere.

Files

• tools/make_replicate_report.py, plus its test (13 tests). Output goes to the darkroom by default, with --also for the repo copy.
• docs/learned/tuned_vs_coact/replicate1/: report.html, merge_gap.json and crowded_check.json for this draw.
• docs/reviews/replicate1-report_2026-09-19.md, the murderboard record, and …-roles/ with all 33 role reports: 3 rounds of 11.
• Three todos, an INDEX row, and the release of the darkroom claim.

Review status. Three blind rounds were run. They did not converge: round 3 found one blocking item and 34 major findings. All were repaired, but the repaired build has had no fourth blind pass. The open items for Tony are in the record.

Depends on. The first draw's merge_gap.json and crowded_check.json are on nets/fair-comparison-report. The architecture drawing is on #660. The generator's defaults point at where those land on main, and this build passed them explicitly.

Darkroom copies:

• the report: <darkroom>/bugarach/2026-09-18-replicate-run-status/report/report.html
• the review rounds: the same report/ folder
• the run's full results: …/results/

🤖 Generated with Claude Code