Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix. Seed 2026-11-09, drawn after earlier rounds today; the platform stratum had no unaudited directory.
The number
0 of 5 meet the bar (the four randomly sampled subjects: 0 of 4, so a Wilson 95% interval of 0% to 49%; the flagship is chosen, not sampled). All five are finished (completeness 3 or more). Across all 139 sampled subjects so far, 24 meet the bar (Wilson 12% to 24%).
Verdicts
| subject | stratum | cor | com | ver | cra | cla | fit | dom |
|---|---|---|---|---|---|---|---|---|
ventures/photo-editing | flagship | 3 | 3 | 2 | 2 | 2 | 2 | 2 |
ventures/world-clock-widget | venture | 2 | 3 | 2 | 3 | 2 | 3 | 3 |
ventures/hardwood-refinishing-bid-calculator | venture | 3 | 3 | 3 | 3 | 2 | 2 | 2 |
research/rank-elicitation | research | 3 | 3 | 2 | 2 | 2 | 1 | 2 |
research/mimic-batch | research | 3 | 3 | 2 | 3 | 3 | 2 | 2 |
- photo-editing. Metrics recomputed with numpy and ffmpeg match. Worst defect:
site/index.htmlstill says the renderers are dry-run placeholders and tells users to run ashootflowcommand thatpyproject.tomldoes not install. Third round with this unchanged. Three of five seeded mutations (exposure target, sharpness half-point, dHash) survived. - world-clock-widget. All 160 UTC offsets match Python
zoneinfo; 320 sunrise/sunset instants agree with an independent NOAA implementation to a median 0.16 minutes. Worst defect: in Reykjavik for about 13 days of midsummer the web port's sunset fell before its sunrise on one local date, giving a daylight length of "-3h -51m". Swift is not installed in the audit environment, so the Swift core was read, not run. Two of three solar mutations (about 2 and 4 minutes) passed all tests. - hardwood-refinishing-bid-calculator. Three jobs recomputed to the cent; 102 tests pass; both seeded bugs caught. Worst defect: screen-and-recoat plus stain is priced with no warning, though a finished floor cannot be stained without a full sand. The README says sand-and-stain costs "several hundred dollars more"; the tool gives $1,416 more.
- rank-elicitation. The paper's numbers reproduce exactly. It is the rearrangement inequality applied to a linear score, cites no prior work, and compares a ranking report against a Brier report that also elicits magnitudes. One mutation (
z*ztoz) survived. - mimic-batch. Results reproduce byte-for-byte. The headline (false-positive rate inflating with batch size) rests on one seed with no interval, and the tests cannot detect the calibration/test leakage the page is about (that mutation survived).
What changed
One auditor-found defect was small and safe: the Reykjavik daylight bug. It now has a fix and a regression test that fails without it (28 pass, 1 fail before; 29 pass after). I re-ran that failing case myself as the spot-check of the audit's claim. The other defects are in the audit files and not fixed here.
Not at the bar
Verification is the most common gap: seeded mutations survived in four of five subjects. The recurring pattern is tests comparing a function to itself or using loose tolerances.