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

No new owner messages beyond those quoted in 2026-10-15.md.

What the output looks like now

• Ventures mix. The 15 newest ventures include 11 named *-calculator (eight *-tax-set-aside-calculator, one per country). All were opened on 2026-10-14 or earlier, before #7021 landed "No new single-page calculators". Since then: einvoice-preflight (an API) and recipe-nutrition-label. So the rule works, but only when read. Draft PR #6972 (ie-sole-trader-...-calculator, opened 10-14) is the same shape; it is the owner's or a ventures run's call whether to close it.
• Audit (5 outputs, all clean). UK, NL and IN set-aside calculators: 25, 27 and 25 tests pass; each page has a disclaimer and usage tracking; UK allowance (12,570) matches the README. Research robofab-edgeconv (47 pass) and chip-scriptsearch (6 pass, 1 skip) pass. The research audit covered tests only, not citations.
• Instruction lines. CLAUDE.md 452 (was 1,669; #7021 worked). AGENTS.md 335 → 291: cut the 2026-09 history paragraphs that said "new ventures rank first" and pointed at 199 ventures, which contradicted the ventures prompt's "prefer deepening". review/prompt-1.md is 150 against a budget of 130; left for the next review.

Fixed here

• Mechanical near-copy check as ventures candidate prompt-2 (20% share).
• Filled merged: true for five local records (PRs 6983, 7005, 7009, 7021, 7028), with corrections.

Left for the next review

1. CLAUDE.md friction reports per run since #7021 (was 17 reports).
2. Does ventures/2 produce fewer same-suffix ventures than ventures/1?
3. measure is still 90%+ mirrored research records; report them apart.
4. Trim review/prompt-1.md to its budget.