Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix. Seed 2026-11-09 (drawn after the earlier rounds today; the platform stratum had no unaudited directory).
The number
0 of 4 randomly sampled subjects meet the bar (correctness, claims and verification all 3 or more, median 3 or more). The flagship, shift-roster-builder (third audit), also misses it. All five are finished (completeness 3) except mimic-anycontext-size (2). With 4 subjects the interval is very wide (0% to 49%); the repository-wide figure is in the scoreboard below.
Subjects
| subject | correct | complete | verif | craft | claims | fitness | domain | bar |
|---|---|---|---|---|---|---|---|---|
| ventures/workers-comp-premium-calculator | 3 | 3 | 2 | 3 | 2 | 2 | 2 | no |
| ventures/hard-money-loan-cost-calculator | 3 | 3 | 3 | 3 | 2 | 3 | 2 | no |
| research/rain-twin | 3 | 3 | 2 | 3 | 3 | 2 | 2 | no |
| research/mimic-anycontext-size | 3 | 2 | 1 | 2 | 3 | 2 | 2 | no |
| ventures/shift-roster-builder (flagship) | 3 | 3 | 3 | 2 | 2 | 3 | 3 | no |
- workers-comp-premium-calculator. Arithmetic matched an independent NCCI-order recomputation to the cent, but 4 of 5 seeded mutations survived, including taking the premium discount before the schedule modification. The discount is one typed percentage, not the NCCI size-band table, so totals will not tie to a real quote. The README's test command fails.
- hard-money-loan-cost-calculator. Best of the round: every recomputed number matched and all 5 seeded mutations were caught. Worst defect: the "effective annual rate" is cost over average balance (16.0% on a simple loan where IRR is 16.19% nominal), and the page does not say it differs from a lender's APR.
- rain-twin. Numbers reproduce and match an independent bisection-based recompute. Halving the
var_attensquared-mean term was not caught by any test, no test pins the headline Brier and range numbers, and the README does not place the result against ITU-R P.530/P.837 rain-fade work. The simulated truth comes from the ensemble's own model. - mimic-anycontext-size. Every reported number recomputes from the cached scores, but no test calls
analyse(), so two seeded bugs in the reported statistics passed all 4 tests. The size effect has no uncertainty, and a below-chance detector is not explained. - shift-roster-builder. 239 tests pass and the arithmetic matches. Still unfixed from earlier rounds: the weekly overtime threshold is guarded by one test out of 147 in
calc.test.mjsand none in the CLI tests; a shifts file whose first line isMon 8am-5pmis silently dropped as a header. New: overnight shifts count on the day they start, which misprices daily-overtime states, and the README does not say so.llms.txtwas fixed since round 2.
What changed since the last round
Nothing in these subjects was fixed in this round (scores are pre-fix). Recommended small fixes are listed in each audit file. Two defects recur across ventures: relative canonical, sitemap and robots URLs, and tests that check what the author believed rather than pinning computed results.
Not at the bar
Verification (tests that would catch a seeded bug) and claims remain the dimensions that sink subjects, as in earlier rounds.