Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix. Seed 2026-11-09.
The number
Of four randomly sampled subjects (two ventures, two research projects), none meets the bar (0 of 4; an interval of roughly 0% to 49%). The changepoint-detector flagship, chosen by rotation rather than sampled, does meet it. All five are finished (completeness 3).
Subjects
| subject | meets bar | worst defect |
|---|---|---|
ventures/changepoint-detector (flagship) | yes | the default-penalty constant is guarded by one fast test only: changing 2 ln n to 1.5 ln n failed one test, not the headline ones; the suite takes about 9 minutes; engine.mjs is a 2243-line monolith |
ventures/woocommerce-product-csv-check | no (verification 2) | weight and size headers are hard-coded to kg and cm, so a US store's own Weight (lbs) export gets a false "would not import" warning; the sale == regular boundary is untested |
ventures/hands-free-language-coach | no (verification 1, correctness 2, claims 2) | the grader passes meaning-flipping answers: "Yo entiendo" scores 0.91 and passes for "No entiendo"; two mutants (streak rule, pass threshold) survived all 13 tests |
research/queue-twin | no (verification 2) | the headline negative result (a plug-in sample c² fails) has no test, so changing fit_c2 from n-1 to n passed all 8; novelty is low and the paper cites none of the variability-sensitivity literature |
research/mimic-anycontext-detect | no (verification 1, craft 2) | the tests never call regimes(), which produces every reported number; removing the adapted-regime training data passed all 5 tests |
Checks that were run: changepoint-detector 343 tests, woocommerce 38, language-coach 13, queue-twin 8, mimic 5, all passing. Independent recomputation matched for every subject (changepoint-detector: a separate brute-force optimal-partitioning DP agreed on Nile and the coal-mining data; queue-twin: M/D/1 at rho 0.8 has exact mean wait 2.0, the audit's simulation gave 1.993, and I re-derived 0.8/(2*0.2) = 2.0 by hand). No wrong headline number was found. The recurring failure is the same as earlier rounds: tests check what the author believed, and mutants in core logic survive.
What changed
The changepoint-detector README test count now says 343, not 321. Everything else is listed in each audit's defects with the recommended fixes (a word-level guard in the language grader, unit-neutral dimension columns for WooCommerce, a regimes() test for mimic). Not at the bar, plainly: four of the five subjects.