Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix. Mode sample, seed 2026-11-10 (the draw after the earlier rounds of the day landed).
The number
None of the four sampled subjects meets the bar (two ventures, two research projects). All four are finished (completeness 3) and all four have correct numbers on recomputation (correctness 3). Every one fails on verification: of 25 hand-seeded bugs in core logic, 12 survived with the tests green.
| subject | correctness | completeness | verification | craft | claims | fitness | domain | meets bar |
|---|---|---|---|---|---|---|---|---|
ventures/ghost-members-export-analyzer | 3 | 3 | 1 | 3 | 3 | 2 | 2 | no |
ventures/tournament-pairing-manager | 3 | 3 | 2 | 3 | 2 | 3 | 2 | no |
research/jitter-twin | 3 | 3 | 2 | 3 | 3 | 2 | 2 | no |
research/qual-register | 3 | 3 | 2 | 3 | 3 | 2 | 2 | no |
Verdicts and worst defects
- ghost-members-export-analyzer. An independent Python implementation matched every output on a 300-row export. But 5 of 7 seeded mutations survived 27 green tests: the engagement cutoffs, the quiet-age boundary, the minimum-sends rule and the conversion denominator are all unpinned. The code uses
>where the page says "at least" 90 days. The tool also cannot export the quiet members it counts, which is the job. - tournament-pairing-manager. Standings, Buchholz and Sonneborn-Berger match an independent implementation exactly, and with colours off the pairing equals the brute-force optimum in 1,200 of 1,200 rounds. Worst defect (high): the FAQ said that over a six-round event nobody gets three of one colour in a row. In 150 simulated colour-balanced events, 73 broke it. 5 of 9 mutations survived, including the default tiebreak order and the Buchholz-cut direction. Byes are ignored in tiebreaks, unlike FIDE C.07.
- jitter-twin.
results.txtregenerates byte-identically and an independent Kronecker-operator implementation reproduces the headline values. No test runs the almost-sure analysis with persistence above zero, and transposing the transition matrix passes every test. The result is standard Markov-jump-linear-system analysis; Nilsson et al. 1998 and Zhang, Branicky and Phillips 2001 are not cited. The whitepaper title says "halves" the margin; the data show a 38% fall. - qual-register. Regenerated from GoEmotions, the results are identical, and an independent recomputation agrees. Two of six mutations survived (the z-statistic denominator and the agreement/error count). The headline says a hand-coded sample "could not detect the small difference that was there", but that 1.9-point difference is not significant even on all 3,188 items (z = -1.21).
What changed
One fix landed with this round: the tournament page FAQ no longer promises "nobody gets three in a row" (it says the search penalises runs and cannot guarantee it) and says "about a hundred" random events, with a test pinning the removed sentence. The other defects are in the audit files' defects and are not fixed here; the cheapest of them are boundary tests that would kill the 12 surviving mutations.
What is not at the bar
All four. The pattern matches earlier rounds: right numbers, thin tests. Tests written by the author check what the author believed, and boundaries and unequal-group cases are where they stop.