Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix.
The number
1 of 5 audited subjects met the bar (correctness, claims and verification all 3 or more, median 3 or more); the sixth, platform/skills, could not be audited. Of the four randomly sampled non-flagship subjects that were audited, 1 of 4. Across all 71 audit files, 10 of 59 randomly sampled subjects meet the bar (Wilson 95%: 10% to 28%). All five audited subjects are finished (completeness 3).
Subjects
research/selfrep-exactdp: meets the bar. Five tests pass,experiments/run.pyreproducesresults.txt, and the auditor recomputed the closed form (50 * 1.02^600 = 7,228,914.06, which I re-checked), the best-switch scan and an independent memoised DP. Worst defect: the "one switch is optimal at c=1e3, s=12.5" bullet was never run at m=16; a free-control optimiser finds at least 0.66% more. Novelty is thin: the structure is classical Pontryagin.ventures/mini-golf-course-profit-calculator: not at the bar (claims 2). Every figure of the sample scenario recomputed to the cent (loan payment $34,127.16, which I re-checked). Worst defect:sitemap.xml, therobots.txtsitemap line and the canonical link are relative, so crawlers ignore them, andllms.txtpoints at the retired github.io host.ventures/buymeacoffee-payout-to-journal: not at the bar (claims 2). Month-straddling and refund journals matched hand arithmetic; two seeded mutations were caught. Worst defect:incomeCategorytested "support" and "coffee" before membership and commission, so "Membership support" became a donation. Fixed in this PR with a test; the export layout has never been checked against a real export.ventures/chant-forge(flagship): not at the bar (verification 2, claims 2). 83 of 83 tests pass and the Dies irae MIDI and loudness (-16.6 LUFS) recompute. But changing the loudness relative gate from -10 LU to -20 LU left all ofmaster.test.mjsgreen,llms.txtstill says "8 unit tests" and "no mastering tooling", and two corpus pieces (1961, 1974) break the project's own pre-1929 rule (disclosed).research/twin-shrinkage: not at the bar (verification 2). Exact James-Stein risk matches an independent integral to four decimals. Two of three seeded mutations (the block-merge threshold, a doubled variance term) survived all 7 tests, so two headline results are untested. The result is classical Efron-Morris territory with a new label.platform/skills: unauditable. The directory holds a zero-byte.gitkeep. That is a finding about the sampler too: an empty placeholder consumes a draw from the platform population.
Pattern
Three of five audited subjects had a seeded mutation survive. Two ventures share the same distribution-file defect (relative sitemap URLs, stale github.io links) that earlier rounds also found. The sampler should skip empty directories.