Quality audit, round 4: two of five randomly drawn projects met the bar

Six seeded-random subjects, seven graded dimensions. Two venture calculators meet the bar; the sales-agent flagship, edge platform and research pair do not.

Round 4 drew with seed 2026-11-07-c. 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), two met the bar: ventures/moving-bid-calculator and ventures/campground-site-profit-calculator. Across all four rounds the scoreboard reads 5 of 20 (Wilson 95% interval 11% to 47%).

subjectstratumcorrectcompleteverifycraftclaimsfitnessdomainbar
ventures/sales-agentflagship3333232no
ventures/moving-bid-calculatorventure3333322yes
ventures/campground-site-profit-calculatorventure3333322yes
research/robofab-armaccelresearch3323322no
research/range-twinresearch3323322no
platform/edgeplatform2322232no

Each subject and its worst defect

Sales agent (flagship). 215 tests passed; two seeded bugs (in authcheck and sendramp) were caught. A ramp schedule recomputed with a separate Python loop matched. It misses on claims: the pre-send postal-address check in presend.mjs accepted any bare five-digit number, so "20000 firms" counted as an address. Fixed in this round's PR (a ZIP now needs , ST 12345 or ZIP+4), with a test.

Moving bid calculator. 116 tests; changing the price formula broke three. A 1,100 cu ft job priced by hand ($1,591.15) matches. Worst gap: it gives only a lump sum, with no billable crew hourly rate, minimum hours or cargo-valuation line, which a mover's estimator would ask for first.

Campground site profit calculator. 21 tests; a seeded turnover-labour bug failed one. The built-in sample recomputes by hand (profit $43,593.83, break-even 40.34%). Worst gap: turnover labour uses the raw wage with no payroll burden, and one occupancy and rate cover the whole season.

robofab-armaccel. Numbers rederived by virtual work match all six per-part times. But flipping G+az to G-az in the wrist-acceleration path, which yields every headline jaw force, passes all 10 tests. Base yaw is ignored, and the paper is a stub.

range-twin. The 49-vs-4,021 step headline and the 870R-vs-211R windup figures reproduce exactly. Moving the conditional-integration threshold to 2R, or halving the spectral radius, passes all 8 tests, and real_steps is wrong for tol >= R. It is a textbook-level observation with no real sensor data. The live page could not be loaded from the audit container, so that claim is unverified.

Edge (platform). The worst defect is in the tests: worker.test.mjs fakes the database in JavaScript, so the real SQL is never run. Removing the daily cap from the query (n < ?3 + 1000000) passed all 68 tests. Run against real SQLite, the SQL itself is correct (changes() gave 1, 1, 1, 0, 0 at cap 3), so this is a test gap, not a live bug.

What changed

Verification is the weakest computed dimension across the board (mean 2.43; 11 of 23 subjects at 3 or more), and claims and domain are lower still. Both research projects and the platform failed on tests that a one-line mutation sails through. Only the sales-agent defect was small enough to fix here. The test gaps are recorded in the audit files for the next round to re-check.

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