Regenerate FINDINGS.md / paper numbers from recomputed artifacts (post-#305)

[openclaw-dv]

PR #305 (recompute-don't-trust) changed published SPS values (e.g. the 3-model ensemble on globalopinionqa moved from ~0.747 rank-6 to ~0.81 rank-1; opinionsqa/subpop ensembles ~0.877/0.879). Hand-maintained numbers in FINDINGS.md, the arXiv draft under docs/papers/, and the Q2 report are now stale/contradicted by the leaderboard.

• Regenerate the findings tables + publish.py::_build_findings narrative from leaderboard-results/ at publish time; CI-diff FINDINGS.md against artifacts so they can't drift.
• Baseline-anchor headline claims (random baseline scores ~0.71–0.76 on the same scale — surface normalized_sps).
• Version the metric (SPS-3 vs SPS-5) and recompute all tables on one scale.
  See gap-analysis P2-1/P2-2/P2-3.

[openclaw-dv]

PR #315 (in CI) implements this. One follow-up it surfaces that is data-work rather than code: the committed SubPOP ensemble still blends subpop_synthpanel_claude-haiku-4-5-20251001_20260411_073013, which the validity filter now excludes as uniform-garbage — so the published 0.879 likely understates a clean 3-model re-blend. Re-running that constituent (Haiku 4.5 on subpop) and regenerating the blend is the cheap fix; the findings pipeline will pick up the corrected number automatically once the artifact lands. The generated findings block flags this caveat automatically until then.

[openclaw-dv]

Shipped in #315. The findings block is now computed from leaderboard-results/ artifacts via the same recompute path that ranks the leaderboard (src/synthbench/findings.py); FINDINGS.md's quantitative sections are script-generated between markers with a CI drift guard (tests/test_findings_drift.py) so narrative and data can't diverge again. Every headline is baseline-anchored against the measured random floors (0.710/0.763/0.757) and labeled with its composite convention.

Headline corrections now published: OpinionsQA ensemble 0.877 (the 0.900 claim retired as non-reproducible), GlobalOpinionQA ensemble 0.813 rank 1, SubPOP 0.879 (with an automatic caveat that one constituent is excluded-as-invalid — clean re-blend tracked in the follow-up comment), ensemble gain restated as +2.7/+4.8/+5.8 pts vs best single at matched n, conditioning asymmetry corrected 2.4×→2.2×. The Q2 report and arXiv draft carry dated errata banners pending editorial rewrite (#258).

Remaining data-work follow-up (in-thread above): re-run the excluded SubPOP Haiku constituent and re-blend.

[openclaw-dv]

Re-blend follow-up completed in #321: clean Haiku constituent (5.0% uniform, SPS 0.809 — in line with the April replications) replaced the validity-excluded run; subpop ensemble recomputed 0.879 → 0.831, still rank 1, verified badge; the findings excluded-constituent caveat is gone. Cost ≈$4.46.