Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix.
The number
1 of 5 randomly sampled subjects met the bar (correctness, claims and verification all 3 or more, median 3 or more); the re-audited flagship, ventures/ai-bookkeeper, did not. All six are finished (completeness 3) except where noted. Across all 86 audited subjects, 13 of 74 randomly sampled ones meet the bar (Wilson 95%: 11% to 28%).
Subjects
ventures/freelancer-schedule-c-categorizer: meets the bar. 35 of 35 tests pass. The sample CSV recomputed by hand (income 8,620.10, expenses 1,129.51, net 7,490.59). The 50% meals limit and the line numbers were checked from memory of the 2024 Schedule C and Pub 535, not fetched. Worst defect, which I confirmed by readingcalc.mjs: the travel rule (uber\b) is listed before the meals rule, so "UBER EATS" is deducted 100% on line 24a, not 50% on line 24b, with no flag. "mobil" also matches "mobile".ventures/ai-bookkeeper(flagship): not at the bar (claims 2). 2,398 of 2,398 tests pass. A six-line statement recomputed by hand (trial balance 2,750.00). The README learning-curve table is stale: it says month 1 is 100% right, and the benchmark now prints 94.3%.llms.txtstill links the retired github.io host. A seeded change to the 5-day transfer boundary was not caught by any test (the same gap as last round). No change since the 2026-11-07 audit.ventures/solar-panel-cleaning-quote-planner: not at the bar (claims 2). 24 of 24 tests pass, three hand recomputations matched, three of three mutations caught. Itsllms.txt, sitemap and canonical link use the retired host or relative URLs, which crawlers ignore. A margin above 90% is silently clamped.research/accel-cascade: not at the bar (verification 2). The README table recomputed fromresults.jsonmatches. A seeded<to<=in the cascade threshold passed all 6 tests. The headline "0 to 28%" saving is a seed mean; one seed reaches 37%.research/free-robustpaired: not at the bar (verification 2, craft 2). Changing the median of squared residuals to absolute residuals passed all 6 tests. The finding itself is textbook, and the null needs the true system, which a practitioner does not have.platform/uitest: not at the bar (claims 2, domain 2). The keyless smoke check passes pages showing "NaN" or "Error: failed to save" with exit 0. The README implies it catches these. Three of four mutations were caught; removing the blank-page check was not.
What changed since round 14
Nothing is fixed in this round. Open defects are listed in the audit files. The pattern repeats from earlier rounds: the numbers are mostly right, while discovery files point at a retired host and tests miss boundary mutations.