Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix. Seed 2026-11-10 (mode sample). No flagship or platform directory was drawn: all were audited within the skip window.
The number
None of the four uniformly sampled subjects meets the bar (two ventures, two research projects; with n = 4 the Wilson 95% interval is about 0% to 49%). After this round the running estimate across all sampled subjects is 25 of 154 (11% to 23%). All four are finished (completeness 3).
| subject | correctness | completeness | verification | craft | claims | fitness | domain | meets bar |
|---|---|---|---|---|---|---|---|---|
| ventures/ghost-kitchen-brand-picker | 3 | 3 | 2 | 3 | 3 | 3 | 2 | no |
| ventures/tire-shop-ticket-margin-calculator | 2 | 3 | 3 | 3 | 3 | 2 | 3 | no |
| research/invthrust-twin | 3 | 3 | 2 | 3 | 3 | 2 | 2 | no |
| research/qual-negint | 3 | 3 | 2 | 3 | 3 | 2 | 2 | no |
Subject by subject
- ghost-kitchen-brand-picker (verification 2). The optimiser is right: an exact-fraction brute force over all subsets matched
analyze()on 300 random instances with 0 mismatches, and the smash-burger example recomputes by hand to $6.86 per order and $5,249.43 a month. But 6 of 10 seeded bugs survived, includingWEEKS_PER_MONTH = 4, which makes every monthly figure 7.7% low with all 25 tests green. Labour is also charged on prep minutes only, not staffed hours. - tire-shop-ticket-margin-calculator (correctness 2). Worst defect of the round:
Tires,$1,200,800,60,100parses as price $1, parts $200, 800 labour minutes, so a thousands comma silently shifts every field. A tire set over $1,000 is the tool's main line, and the output looks plausible. Its arithmetic otherwise recomputes (average ticket net 213.6025 matches Python); 5 of 6 seeded bugs were caught, and the survivor removes the card-fee factor from the discount's lost net, because the discount test runs at 0% fee. - research/invthrust-twin (verification 2). The numbers reproduce byte for byte and the 0.0460 overshoot matches the closed form exp(-pi*0.7/sqrt(1-0.49)) = 0.04599. But no test covers the linear inverse when the assumed exponent differs from 2, the project's headline regime: changing
n_hat * gto2 * gleaves all 7 tests green. The 1.5 to 2.5 exponent range is assumed, not sourced, and the nearest prior work (PX4's thrust-curve blend) is not cited. A small sensitivity sweep, below the programme's PhD bar. - research/qual-negint (verification 2). Reproduces byte for byte (45 s); the Wilson constants in the tests match a hand formula. Two seeded bugs survived:
plus_twowith(x+1, n+1)and a 0.5 stratum weight, which feeds every coverage number. The result is Brown, Cai and DasGupta (2001) applied to one estimator on one population; the weighted-Wilson literature is not cited.
What binds
Verification again: three of four fail on it alone, and the fourth fails on a wrong number. The pattern from earlier rounds holds: the outputs are mostly right, and the tests would not notice if they stopped being right.
What changed
One fix in this PR: ghost-kitchen-brand-picker gets a test pinning WEEKS_PER_MONTH = 52/12, the per-order profit and the example's $9,026.67 optimum. With the constant mutated to 4 the new test fails and the suite was otherwise green. The other defects (the comma parsing, the discount test at a non-zero fee, the two qual-negint tests, the invthrust-twin exponent test) are recorded in the audit files and not yet fixed; a later round re-audits.