Drift report, 2026-10-15 (second review)

Short follow-up to 2026-10-15.md, written by the first scheduled review. No new owner messages were available beyond those quoted there.

Since the last report

• Ventures mix. One new venture directory landed (einvoice-preflight, a self-hostable validation API with Docker and a Render deploy file). It is not a calculator. No new *-calculator directory appeared in git history. This matches "no more single-page calculators".
• Self-deploy. platform/deploy/ landed: a one-click page that deploys a static venture to the visitor's own GitHub Pages, with secrets sealed in the browser. This is the owner's business model (quoted in goals.md). Its 20 tests pass. How many ventures use it is not yet measured.
• Instruction lines on main: CLAUDE.md 1669, DIRECTION.md 987, AGENTS.md 335, goals.md 252, review/prompt-1.md 144. Flat or up against the last report, because the prune is not merged.
• Prune in flight: draft work in PR #7021 cuts CLAUDE.md from 1,662 to 451 lines. This review did not duplicate it; it only checks the result.

Evidence the rules still cost time

CLAUDE.md has 15 friction reports, the most of any file. The recent ones are the checklist's test commands exceeding the tool timeout, the shallow clone, and pip install pytest landing in the wrong interpreter. All sit in the incident-history sections that #7021 removes.

Fixed here

• prompts/review/prompt-1.md step 1: skip records mirrored from the upstream research repo when filling merged (167 of 198 records, so merged can never be filled for them). One friction report.
• Filled merged: true for four local records (PRs 6987, 6994, 7003, 7020), with correction entries.

Left for the next review

1. Check CLAUDE.md line count and its friction reports per run after #7021.
2. Count ventures with a working self-deploy page.
3. Audit five recent outputs properly. This run checked two (the deploy page and einvoice-preflight: 20 and 251 tests pass, disclaimers present) and no research output; research is 94% of records, so it is the gap.
4. measure is dominated by mirrored research records (185 of 198), which makes every other track incomparable. Consider reporting them separately.