Review, 2026-10-17

Window: the 40 runs since the review of 2026-10-16. Source: prompts/runs/ (713 records), promptctl.py measure, derive.py fetch/build.

What happened

• Success: research/0 97% (391 runs), research/1 100% (36), ventures/1 86% (97), ventures/2 83% (30), process/1 70% (57, up from 56% on 2026-10-16), explore 90%, redteam 95%. Leaked work: none new (7 on old records).
• Open PRs: 20, of which 12 are adopt_stranded.py drafts from branches dated 2026-09-23/24 (#7544-#7553, #7471, #7472), each a comparison or guide page with no CI run yet. They are cheap to land or close; the next ventures or process run should take them.

What is off

1. Starting-message table was noise. 49 of 60 rows were single runs. Fixed (E30).
2. CLAUDE.md is still the top friction file (141 reports). The newest ones say the product checklist does not say which subset applies to a research-only PR. Proposal in the backlog is already P017 (path-scoped instructions); not repeated here.
3. Ventures mix: recent ventures PRs are mostly page additions and features on existing calculators (league-schedule-generator, shift-roster, review-reply-desk), which matches the owner's "SEO and usage first" direction better than new single-page calculators.

Proposals for the owner

None new.