Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix.
The number
0 of 4 randomly sampled subjects meet the bar; the chosen flagship does. Across all rounds: 19 of 111 sampled subjects meet it (Wilson 95% interval 11% to 25%); 105 of 111 are finished.
Subjects
| subject | verdict | worst defect |
|---|---|---|
ventures/cake-pricing-calculator | finished, below bar (claims 2, domain 2) | llms.txt linked the retired GitHub Pages host. Independent exact-fraction recomputation of 3 cakes matched; no fixed card fee, sales tax or waste factor. |
ventures/mosquito-control-program-pricing-calculator | finished, below bar (claims 2) | A blank or non-numeric input silently falls back to the example default, so a job can be priced on a placeholder without warning. llms.txt linked the retired host. Seven outputs matched Python to the cent. |
research/inertia-twin | finished, below bar (fitness 1) | The headline is a textbook identity measured against a twin defined to have no inertia, true by construction. Also a friction-sign bug on the first slip step that a seeded mutation exposed as untested. |
research/replay-memory | finished, below bar (claims 1, correctness 2) | The "exact" 2x2 closure resamples the stored subspace at every replay; at d=12, r=4, step 6 it gives 0.0061 against 0.0071 from a 300k-run simulation. The paper calls the gap Monte-Carlo noise, which it is not. |
ventures/wall-framing-takeoff (flagship, third audit) | meets the bar, 3/3/3/2/3/3/3 | Header spans and jack counts are unchecked unless the user supplies a header table. A mutation to the cap-plate joint window was caught by only 1 of 134 tests. |
What changed
Both calculators' llms.txt now point at marisstella.app (PR below). The research defects and the framing gaps are recorded in the audit files, not fixed here; a later round re-audits.