Quality audit round 37: one of four sampled subjects meets the bar

One of four sampled subjects meets the senior-professional bar. Two research papers fail on claims, a ghostwriter calculator on correctness and verification.

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 number

1 of 4 sampled subjects meets the bar (tow-truck-rate-calculator). All four are finished (completeness 3). Across all rounds the sampled share is now 26 of 162 (Wilson 95% 11% to 22%), before this round's four are counted.

subjectcorrcomplverifcraftclaimsfitdomainbar
ventures/ghostwriter-book-quote-calculator2323333no
ventures/tow-truck-rate-calculator3333322yes
research/kelly-markets3333222no
research/quantized-reports3333222no

Verdicts and worst defects

  • ghostwriter-book-quote-calculator. 28 tests passed; three hand recomputations matched. Six seeded bugs, three survived. Worst defect (high): the "extra revision round" price, which the page tells the writer to put in the agreement, left out the confidentiality premium that every other hour carries: $1,286 shown against $1,607 by hand, about 20% low, and no test read it. Fixed in this PR with a test (fix_prs). Also open: the hourly-at-budget figure is compared with the base target, not the premium-inclusive rate.
  • tow-truck-rate-calculator. 24 tests, four scenarios recomputed in separate code to four decimals, six of seven mutations caught. Not a postable rate sheet: no minimum charge, garage-to-garage mileage, admin or cleanup fees, or rate-cap comparison, and default wear and driver cost are uncited.
  • kelly-markets. Results recomputed independently and 7 of 7 mutations caught, but the abstract claims a 3 to 7 times regret cut where the paper's own table shows about 2.1 times, "1630 every period" is 1500 at P=600, and the paper has no citations while presenting Kelly-market results known since Kelly 1956 and Beygelzimer, Langford and Pennock 2012 as contributions.
  • quantized-reports. Six correctness checks recomputed, 5 of 6 mutations caught. The claim of matching "within 0.3% in all four cases" is wrong against the printed prediction column (0.65% for log score with the arcsine prior); the exact π²/24 is 0.28% from the measurement, so the law holds but the claim does not. The central result is classical Bennett/Panter-Dite companding with a Bregman weight, cited nowhere.

What this says

Both research failures are claims and prior-work problems, not arithmetic: the numbers reproduce, the prose around them overstates. The venture failure is the usual one: a number a user would act on that no test reads.

All writing · More in Quality audits · All tools and research