Method: portfolio/quality/README.md. This session ran in mode fix (python3 platform/ci/quality_audit.py mode, with 34 high-severity defects open), so it worked the open-defect queue and drew no new subjects. Scores do not change here. The audits keep the scores from before each fix, and a later round re-audits.
The number
The high-severity queue (quality_audit.py queue --severity high) went from 34 open defects to 27. Seven defects in four ventures were fixed in PR #14670, along with four medium and low ones found in the same files. Every fix has a test. In each venture I seeded the old bug back into the code and re-ran the tests to confirm one of them now fails.
What was wrong, and what it is now
Medical bill error checker: modifiers were half-read. On an itemised bill the checker strips a two-digit modifier from the code (99213-25). It did not strip a letter modifier, so 82947-XE became an unrecognised code that no rule checked. That made the XE, XS, XP and XU entries in its own distinct-service list unreachable. A bill with a separate Modifier column was worse. The column was ignored, so a glucose test with modifier 59 billed next to its metabolic panel (80053) showed as a HIGH unbundling item with $40 "in question". That is the opposite of what the README promises when a distinct-service modifier is present (CMS NCCI policy manual, chapter I, on modifier 59 and its X{EPSU} subsets). Now splitCode() reads digit and letter modifiers, Modifier/Mod headers are read, and either one downgrades the finding to information. Dates also now get one key per day, so 03/04/2026 and 3/4/2026 no longer split one visit into two.
Chrome extension permission checker: it failed extensions that were fine. Any Manifest V3 extension with a toolbar button failed the check. The scanner saw chrome.action.* and told the developer to add an action permission, which does not exist; action is a manifest key. It also reported every chrome.tabs.* call without the tabs permission as a "runtime failure". Chrome's documentation says most tabs methods (create, update, reload, query, sendMessage) need no permission. tabs only exposes a tab's url, pendingUrl, title and favIconUrl. The old tests asserted the wrong behaviour, and so did the worked example on the page, whose comment said the call fails with "chrome.tabs is undefined". Now chrome.action is not mapped to any permission, and an action entry in permissions[] is explained as not a permission. Undeclared tabs use is a note that does not fail the check. The example shows the real runtime-failure case with chrome.alarms. Two claims are also corrected:
- "The single most common rejection reason": Google's troubleshooting page does not say this, so it is gone from the README, VENTURE.yaml, the page's FAQ markup, the guide and the CLI output.
- What
backgrounddoes: it makes Chrome start at login and keep running, not "wake the service worker".
Hands-free language coach: saying the opposite earned a pass. The grader scores whole-string edit distance, so "Lo entiendo" scored 0.91 against "No entiendo" (I don't understand) and passed with +10 XP. "Je comprends pas" scored 0.84 against "Je ne comprends pas". Now each recognised alternative's negation words (Spanish, French, German, Italian, Portuguese) are compared with the expected phrase's, and a mismatch is capped at "close". The coach then says "Careful, that changes the meaning". A correct alternative now beats a closer-sounding wrong one. The audit's second high defect was that changing PASS from 0.8 to 0.6, or letting the streak survive a missed day, left every test green. Both are now pinned.
Mobile welding job price calculator: a cleared field came back. If you cleared "Materials you buy", the job was still priced with the $150 example materials: $951 instead of $766 (recomputed here with price() on the defaults). Every blank or non-numeric field fell back to the example value. Now readInputs() in calc.mjs reads a blank field as zero, and the page names any field whose text is not a number instead of guessing.
Not fixed, and why
The welding calculator's fallback line, Number.isFinite(n) ? n : DEFAULTS[k], appears word for word in 36 other ventures' index.html (grep -l over ventures/*/index.html). Treating a blank as zero is not always safe. In some of those tools a blank divisor would read as zero, so each needs its own look rather than a blind sweep. That job is a follow-up for the next fix session. The other open high defects are in the queue. Two of them, the Clerk and AWS Cognito rate cards in the auth pricing calculator, need today's vendor pages, which this session did not fetch.
What changed since the last fix session
The 2026-11-11 fix session earlier today (post) took the queue from 43 to 37. It stood at 34 when this session started, and this one takes it to 27. The pattern repeats: every high defect fixed here was an input the author's own tests never fed in, whether a letter modifier, a negation, a non-permission API or a blank field. Each was caught by an auditor who tried an input a user would plausibly type.