Quality audit round 39: one of four sampled subjects meets the bar

One of four sampled subjects meets the senior-professional bar: a gig-driver earnings analyzer. A trampoline-park model and two research twins miss.

Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix. Mode sample, seed 2026-11-10 (no flagship or platform subject was drawn: all were audited recently).

The number

1 of 4 sampled subjects meets the bar. Across all sampled rounds, 28 of 170 meet it (Wilson 95% interval 12% to 23%). Four subjects is too few to say anything about this round alone.

subjectcorrectnesscompletenessverificationcraftclaimsfitnessdomainbar
ventures/gig-driver-earnings-report-analyzer3333323yes
ventures/trampoline-park-profit-calculator3323222no
research/lag-twin3323222no
research/ransac-twin4323212no

Verdicts and worst defects

gig-driver-earnings-report-analyzer (meets the bar). 33 of 33 tests pass; a two-trip file recomputed by hand (net 6.60, 7.92 per hour, 0.44 per mile) matches the tool, and I re-ran it. 3 of 4 seeded mutations caught; the survivor is MIN_TRIPS at calc.mjs:232. Worst defect: the page reads any semicolon-delimited file as decimal-comma (index.html:151), so 12.50;3.2 becomes pay 1250 and miles 32 with no warning. A separate Time column is also ignored.

trampoline-park-profit-calculator (fails on verification and claims). 24 tests pass and the sample P&L reproduces to the cent in an independent script (profit 9,311.92, payback 64.43 months). But 3 of 5 seeded mutations survived, including the loan term in the 10%/20%-fewer-visits payback shown on the page. The $250k to $800k build cost and "lenders want under five years" have no source, and the model has no manager payroll or DSCR.

lag-twin (fails on verification and claims). All headline numbers reproduce with an independent root finder. Worst defect: the whitepaper said equality in the 1/(3τ) speed cap holds "iff a triple pole". A one-liner with numpy.roots shows ζ=1.5 and ζ=2 also reach 0.3333, so that was false. Fixed in this PR with a test and a corrected sentence; the audit scores are before the fix. Two of four non-equivalent mutations survive, and the result is a textbook Routh-condition example.

ransac-twin (fails on verification and claims). The exact-arithmetic recomputation matches (failure 0.140776 at K=17, 99% budget 458 for κ=4). But ransac_line, which produces the end-to-end headline rates, has no test: two seeded bugs there left all 6 tests green. The only citation is Fischler and Bolles; adaptive termination work is absent, and "K sized on the mean inlier fraction under-covers" is Jensen's inequality, so the result is not new as framed.

What is not at the bar

Three of four miss, all with verification at 2: tests passed while seeded changes to core logic also passed. The pattern holds across the corpus (verification at 3 or more in 82 of 184 audits). Open defects not fixed here are listed in each audit file.

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