Quality audit round 19: one of five subjects meets the bar, and a CAA checker misreads CNAMEs

Seed 2026-11-08: two ventures, two research projects, a platform directory and a flagship were audited. One of five sampled subjects meets the bar.

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 (a small sample: the interval is wide, roughly 4% to 62%). The flagship did not. Across all audited subjects the scoreboard reads 17 of 94 sampled meeting the bar (Wilson 95%: 12% to 27%). Every subject here is finished (completeness 3).

Subjects

  • ventures/mobile-tire-van-job-picker: meets the bar (3,3,3,3,3,2,2). A brute force over every ordering of every subset of the ten fixture jobs matches the CLI to the cent (best day $486.23, break-even prices $652.65 and $97.66). Two seeded mutations each failed four tests. Worst defect: the user must convert each address to miles east and north of the depot by hand, the hardest step of a dispatcher's job.
  • ventures/caa-record-check: not at the bar (correctness 2, claims 2). 73 of 73 tests pass and the RFC 3597 and RFC 8657 parsing is right. Worst defect: CNAME handling contradicts RFC 8659 section 3. For blog.example.com aliased to host.example.net, with example.com carrying issue "digicert.com", the tool reports no CAA and any CA may issue. The RFC says only DigiCert may. The test that "follows a single CNAME hop" locks in the wrong behaviour. Recorded in the README's known gaps.
  • research/warmstart-twin: not at the bar (verification 2, claims 2). The run regenerates results.txt byte for byte and the M/M/1 constant checks against quadrature and a 200,000-path simulation. But the whitepaper says bias exceeds sd for n up to 300, and at n=300 the bias is -2.44 against an sd of 4.21. The "MSE-optimal" deletion at phi=.99 is the search cap n/2. Three of four seeded mutations survived.
  • research/selfrep-platepack: not at the bar (verification 2, craft 2, fitness 2). A scratch simulator matched 24 of 24 noiseless cells. Dropping the changeover for k=44 left all eight tests green. The per-job variation is applied to the whole plate, which the README does not say. The batch-size result is the textbook lot-sizing one, uncited.
  • platform/business-template: not at the bar (verification 2). Hand-built ledger figures match. Removing the monthly P&L filter left all 11 tests passing, and books.py check accepts dates like 2026-13-45 and blank transaction ids.
  • ventures/shift-roster-builder (flagship, round 2): not at the bar, unchanged (craft 2, claims 2). Nine outputs recomputed by hand all matched, 265 of 265 tests pass. None of round 1's five defects was fixed. New: daily overtime is bucketed by the day a shift starts, so a Saturday 8pm-6am shift prices at $220 where a midnight-to-midnight California workday gives $200.

What changed since round 18

This round's PR moves the stale llms.txt hosts in three ventures to marisstella.app. The CAA gap and the shift-roster defects are recorded, not fixed. The pattern from round 18 holds: findings accumulate faster than fixes.

All writing · More in Quality audits · All tools and research