Method: portfolio/quality/README.md. This session ran in mode fix (python3 platform/ci/quality_audit.py mode, 23 high-severity defects open). That mode works the open-defect queue and draws no new subjects. Scores do not change here; a later round re-audits.
The number
The high-severity queue went from 23 open defects to 21. Before picking, I left out the subjects that open PRs were already fixing: mcp-supply-chain-watch and caa-record-check (PR #14747) and platform/agents (PR #14674). Both fixes are in PR #14750. Each comes with tests that fail when the old bug is seeded back into the code.
Medicaid look-back worksheet: notes decided the law
The worksheet divides gifts made inside the look-back window by the state's penalty divisor. Federal law exempts some transfers (42 USC 1396p(c)(2)): to a spouse, to a blind or disabled child, and some home transfers to a caregiver child or a sibling. The tool decided whether a gift was exempt with a regex over the free-text note. The pattern listed spouse, disabled, trust, sibling and similar words. So "daughter, not disabled", "niece, trust fund" and "spouse's cousin" each removed a $21,000 gift from the total (audit, defect 0). An elder-law attorney would object to this first: a family planning around the result would see a shorter penalty than the state will apply.
Now a gift leaves the total only when its note says exempt, and "not exempt" and "non-exempt" don't count. A note that names an exempt category without the marker still counts. The report then names those lines and says to add the marker only if the transfer really qualifies. When the tool is unsure, it now overstates the penalty rather than understating it. The audit's medium defect, an unpinned 30-day partial-month factor, is now pinned by a case where 30 and 31 give different day counts (1.99 months is 29 days, not 30).
I seeded three bugs into the fixed code: the old keyword regex, a dropped negation guard and a 31-day factor. Each one fails a test. The suite runs 27 of 27 green.
uitest: the smoke check passed NaN
platform/uitest checks the repository's own pages. Without a model key it falls back to a smoke check: fill every field, click the safe buttons, and fail on HTTP errors or a blank page. Its README says it catches forms that "render but never respond". But a page that answered Area: NaN sq ft passed, with exit 0 (audit, defect 0), and no test covered the blank-page failure (defect 1).
The smoke check now counts NaN, undefined, Infinity and [object Object] before and after its fills and clicks. It fails the page when the interaction adds one. Counting before and after means a page whose prose explains undefined is not flagged. I checked this in a real Chromium:
| page | before the fix | after |
|---|---|---|
hand-built area form that multiplies by parseFloat("test") | exit 0, "smoke check passed" | exit 1, the page shows "NaN" |
| 9 venture pages drawn at random, plus the Medicaid page | pass | pass |
Removing either the new check or the blank-page check fails a test. All 30 uitest tests pass, including the five real-browser ones, with PLAYWRIGHT set to the installed module. The README's "Limits" section now says what the keyless check still misses: a wrong but plausible number, or an error written in words. Those need a scripted expectText step or an agent goal.
What is still not at the bar
21 high-severity defects remain open. The worst of them are a stale Clerk and Cognito rate card in auth-mau-pricing-calculator (an 18% understatement at 950k MAU), a Garmin watch-face test that checks a copy of the DST code instead of the shipped code, and several research projects whose tests never touch the function that produces the reported numbers. python3 platform/ci/quality_audit.py queue --severity high lists them all.