Round 3 drew with seed 2026-11-07-b (rounds 1 and 2 used the bare date; a parallel round 2 landed while this one ran, and drew ai-bookkeeper too, so this round's independent ai-bookkeeper audit is kept at docs/reports/2026-11-07-ai-bookkeeper-second-audit.json and not double-counted in the scoreboard). The method and rubric are in portfolio/quality/README.md; each subject's evidence is in portfolio/quality/audits/.
The number
Of the five randomly sampled subjects (not the flagship), one met the bar: research/precision-scores. The flagship, ventures/ai-bookkeeper, did not. Counting this round's five sampled subjects with the earlier ones, the scoreboard reads 3 of 15 (Wilson 95% interval 7% to 45%).
| subject | stratum | correct | complete | verify | craft | claims | fitness | domain | bar |
|---|---|---|---|---|---|---|---|---|---|
ventures/ai-bookkeeper | flagship | 3 | 3 | 3 | 3 | 2 | 2 | 3 | no |
ventures/lash-studio-profit-calculator | venture | 3 | 3 | 2 | 3 | 2 | 2 | 2 | no |
ventures/amazon-seller-profit-analyzer | venture | 2 | 3 | 2 | 3 | 2 | 2 | 2 | no |
research/precision-scores | research | 4 | 3 | 3 | 3 | 3 | 2 | 3 | yes |
research/accel-kwssparse | research | 2 | 3 | 2 | 2 | 2 | 2 | 2 | no |
platform/audio | platform | 3 | 2 | 2 | 3 | 2 | 3 | 3 | no |
Each subject and its worst defect
AI Bookkeeper (flagship). 2,371 tests pass. A seeded bug (period-end > changed to >= in ledger.mjs) broke 12 of them. Self-employment tax on $100,000 recomputes to $14,129.55 (92,350 × 15.3%). It misses on claims: llms.txt said the product "does not do payroll" while payroll.mjs ships Form 941/940 worksheets, and the README promised payees "post themselves from the second month", where the author's own benchmark shows 7.2% in month 2 and about 43% in month 4. Both claims were corrected in this round's PR. Not fixed: 96 cells of the 10-, 15- and 20-year MACRS tables differ from IRS Pub. 946 by 0.01 to 0.05 points, and no reader has been tried on a real bank export.
Lash studio profit calculator. The seeded bug (monthly profit as weekly × 4 instead of × 52/12) passed all 23 tests: the monthly figure, weekly figures and CSV have no assertions. Client lifetime uses only the first set and first fill type. llms.txt points at retired github.io addresses.
Amazon seller profit analyzer. The worst defect is a wrong number: "referral fees Amazon kept" ignores quantity, so a partial refund of a 2-unit order shows $3.96 where the correct figure is $0.66. The waterfall also omits the regulatory fee, so its lines are off by $0.04 from "Paid out", and a seeded bug survived all 23 tests.
Precision scores (met the bar). Recomputations agreed (the efficiency loss at condition number 100 is 4κ/(1+κ)² = 0.0392), and all six seeded mutations were caught. It is not new: the bound is the classical Watson / Bloomfield-Watson result, and the write-up cites nothing. The finite-sample figures come from one covariance with no standard errors.
Accel-kwssparse. Correctness is capped at 2 because the 3.3 GB feature set is not available, so the accuracies were not reproduced. Of three seeded mutations only one was caught; no test pins an energy value. The ternary and binary failure comes from an untuned recipe on a weak baseline, and is expected from the literature.
Audio (platform). A seeded +0.8 dB gain bug in master() passed all 24 tests, because the loudness test allows ±1 LU and the output already lands 0.3 to 0.4 LU under target. Silent input is accepted without error, and no workflow runs test_narrate.py.
What changed since earlier rounds
Same share meeting the bar, and the same pattern: arithmetic is mostly right and what fails is claims and verification. A deliberately planted bug survived the suite in 4 of this round's 6 subjects (lash studio, amazon seller, accel-kwssparse in part, audio), against 3 of 6 in round 1. Re-audits are separate files. Not at the bar: five of these six, the flagship included. Nothing here raises a round 1 score; re-audits are separate files.