Method and rubric: portfolio/quality/README.md. Evidence per subject: portfolio/quality/audits/. Scores are before any fix.
The number
Of four randomly sampled subjects (two ventures, two research projects), none meets the bar (0 of 4; the interval is very wide, roughly 0% to 49%). The flagship, ai-bookkeeper, does not either. Across all rounds, 22 of 119 sampled subjects meet the bar (Wilson 12% to 26%), so this round sits at the low end of the usual range. All five subjects are finished (completeness 3).
Subjects
| subject | meets bar | worst defect |
|---|---|---|
ventures/window-tinting-bid-calculator | no (claims 2) | a blank numeric input silently becomes 0: a blank installer rate gave $284.11 instead of $370.27; the "how the trade prices" claim is unsourced |
ventures/gutter-install-bid-calculator | no (claims 2) | no crew-size input, so a one-wage crew-hours figure undercosts multi-person crews; no waste line |
research/qual-regression | no (verification 2) | removing the 1/pi weight in the DSL influence rows corrupts standard errors and every test still passes |
research/merge-law | no (claims 2) | the paper cites nothing; the result is relaxed parallel projection (Cimmino, block Kaczmarz) in task-arithmetic wording, checked only on its own toy model |
ventures/ai-bookkeeper (flagship) | no (claims 2) | the two defects from the previous audits were not fixed; a new seeded mutation of the $400 self-employment threshold (< 40000 to <= 40000) passes all tests |
Checks that were run: window-tinting 107 tests, gutter-install 122, qual-regression 5, merge-law 7, ai-bookkeeper 2405, all passing. Hand or independent recomputation matched for every subject (Schedule SE on $150k of W-2 wages: 5,884.89, matching the form). Mutations: merge-law caught 3 of 3, qual-regression 2 of 3, window-tinting and gutter-install caught what was seeded, ai-bookkeeper missed the new one.
What is not at the bar
The numbers are right; the weakness is evidence around them. Four of five fail on claims: unsourced trade rates, a stale README benchmark table, a paper with no related work. The fifth fails on verification, where tests pin outputs but not the standard errors that are the headline. The ai-bookkeeper flagship has now been audited three times and the same documentation defects remain open.
What changed since the last round
One small fix ships with this round: the llms.txt files of window-tinting, gutter-install and ai-bookkeeper pointed at the retired github.io address and now point at marisstella.app. Other defects are in the audit files for the next round, which will re-audit rather than raise scores for fixes.