The method and rubric are in portfolio/quality/README.md. The evidence is in the four 2026-11-13-* files with seed 2026-11-13-g in portfolio/quality/audits/. Each file lists every check and every hand-seeded bug. All scores were set before any fix.
This round ran in mode sample. Rounds 55 to 60 had already used seeds 2026-11-13 to -f, so this one took -g. None of the four subjects had been audited before. No flagship or platform subject was drawn, because each had been audited too recently.
The number
0 of 4 sampled subjects meet the bar. Across every sampled audit, 31 of 258 meet it (Wilson 95% interval 9% to 17%). All four are finished (completeness 3). All four miss on verification, as in rounds 59 and 60: in every one, at least one seeded bug in core logic got past the tests.
| subject | correctness | completeness | verification | craft | claims | fitness | domain | bar |
|---|---|---|---|---|---|---|---|---|
| ventures/amusement-route-location-calculator | 3 | 3 | 2 | 3 | 3 | 2 | 2 | no |
| ventures/uk-savings-interest-tax-calculator | 2 | 3 | 2 | 3 | 2 | 3 | 2 | no |
| research/garment-twonodeprop | 3 | 3 | 2 | 3 | 3 | 1 | 2 | no |
| research/eda-rankpayoff | 3 | 3 | 2 | 3 | 2 | 2 | 2 | no |
| subject | seeded bugs caught | correctness checks passed |
|---|---|---|
| amusement-route-location-calculator | 2 of 4 | 4 of 5 |
| uk-savings-interest-tax-calculator | 1 of 4 | 3 of 6 |
| garment-twonodeprop | 2 of 4 | 4 of 4 |
| eda-rankpayoff | 1 of 4 | 5 of 6 |
The tests caught 6 of 16 seeded bugs (round 60: 7 of 20). Both research projects reproduce byte for byte: garment-twonodeprop's results.txt and figure in 27 seconds, eda-rankpayoff's in under a minute. As in round 60, the research numbers are right and the research framing is thin; the products get the main arithmetic right and an edge wrong.
I re-ran two of the auditors' checks myself. The amusement-route engine prints 310.08 / 103.36 / 66.90% / 81.72 for the sample "Corner bar", the same as the auditor's separate script. I worked the UK case by hand (below) and got the auditor's £5,800.
Verdicts and worst defects
ventures/uk-savings-interest-tax-calculator (correctness 2, claims 2)
The reviewer is a UK chartered tax adviser who checks HMRC P800 and Simple Assessment savings figures. Every 2026/27 rate and threshold matches GOV.UK's "Income Tax rates and allowances" table. Both of GOV.UK's worked examples come out right (£16,000 wages and £200 interest: £0; a basic-rate saver with £1,300 interest: £60), and so does a band-edge case (£49,000 + £1,500: £246, where the savings allowance drops to £500).
The worst defect (high, logic error): above £100,000 the personal allowance falls by £1 for every £2 of income, and savings interest counts as income. The tool taxed the interest itself but left out the extra tax on wages that the lost allowance causes. With £100,000 of wages and £10,000 of interest:
- with the interest, the allowance is £7,570, taxable income £102,430, tax £33,432 less £200 of savings-allowance relief: £33,232;
- without it, the allowance is the full £12,570, taxable income £87,430, tax £27,432;
- the difference, which is what HMRC will collect, is £5,800. The tool said £3,800.
The page said it covered "the £100,000 taper". The auditor's grid found 1,980 of 7,865 income/interest pairs wrong, by up to £5,028.
Fixed in this PR: the engine adds the lost-allowance tax and shows it as its own row; the tax-free figure is whole pounds (it showed £500.01 for an income where nothing is tax-free); the "What to do" list says interest over £10,000 needs a Self Assessment return; the ISA advice names the £20,000 yearly limit; and the "written from memory" notes are gone now that the rates have been checked. New tests pin hand-computed figures, and all three seeded bugs that had survived (the taper divisor, the 125,140 threshold, the savings-allowance band edge) now fail the suite.
ventures/amusement-route-location-calculator (verification 2, fitness 2)
The reviewer is a coin-op route owner who sets location commissions. Three sample locations recomputed by hand, with straight-line machine wear, match the engine and the guide's table to the cent.
The worst defect (medium, logic error): when prizes, card fees and commission take all of the gross, there is no break-even take, but the note said "It needs $0.00 a week per machine". Fixed in this PR, with a test; two seeded bugs that had survived (the "meets target" comparison and the break-even guard) now fail the suite.
Still open (README "Known gaps"): commission is taken on gross only, while many operators split after prize cost or fees; each stop is charged a full round trip, which overcharges chained routes.
research/eda-rankpayoff (claims 2, verification 2)
The reviewer works on multi-fidelity design-space exploration. All 14 named-screen rank correlations match scipy's spearmanr and kendalltau to three decimals. The auditor computed the capture probabilities exactly (a hypergeometric sum) instead of by Monte Carlo; the largest gap is 0.011, and the headline holds (0.614 against the reported 0.598).
The worst defects (medium): the README calls this "a new question" but cites none of the neural-architecture-search papers that already compare predictors' rank correlation with their top-k payoff (White et al. and Ning et al., NeurIPS 2021; Abdelfattah et al., ICLR 2021), and tests no top-weighted measure as the alternative. The figure was titled "Rank agreement does not predict screen payoff" while plotting k = 3, where the correlation is +0.221 with a bootstrap interval of 0.068 to 0.360 that excludes zero.
In this PR: a truthful figure title; the README no longer calls a same-seed match "independent code"; a "Known gaps" section with the prior work and the tie-breaking issue; tests that kill the two surviving mutants (Kendall tau-b swapped for tau-a, and the tie-break).
research/garment-twonodeprop (fitness 1, verification 2)
The reviewer is a building-thermal-comfort researcher. An independent reimplementation of both gates and the tuning loop reproduces every headline cell; grids 4 to 10 times finer leave the negative result unchanged.
The worst defect (medium, research design): the README leaves "with a per-switch cost proportional control might help" as the open question. The upstream project's switch-penalty sweep answers it in seconds: at 0.01 C per switch, proportional control wins all four "ours" cells at lags 0 and 10. That near-on/off control wins when switching is free is textbook control theory, and none is cited. The project is one 49-line module, well short of a paper. In this PR: "Known gaps" in the README, and two tests that pin the chosen band and the full transmittance range, which two seeded bugs had left green.
What changed since round 60
Same shape, new subjects: 0 of 4 again, and verification binds in every one. Of the 16 mutants seeded this round, 10 survived; the fixes in this PR make 9 of them fail, and a later round will re-audit rather than raise any score now. The pattern of rounds 59 to 61 holds: research projects reproduce their numbers and overstate their novelty, and calculators get the common case right and the edge case (a taper, a zero denominator, a market restriction) wrong, with no test at the edge.