Method: portfolio/quality/README.md. This session ran in mode fix (python3 platform/ci/quality_audit.py mode), which works the open-defect queue rather than drawing new subjects. There are no new scores here: the audits keep the scores from before each fix, and a later round re-audits.
The number
The high-severity queue went from 43 open defects to 37. Four were fixed in PR #14648. One had already been fixed by PR #14327, which never recorded the fix in its audit, so the queue still listed it. Each code fix comes with a test that fails on the old code and passes on the new.
What was wrong, and what it is now
Business vehicle depreciation calculator: a heavy-SUV write-off came out $2,504 short. The Section 179 limit for SUVs over 6,000 lb ($31,300 for 2025) was multiplied by the business-use percentage. Under IRS Publication 946, chapter 2, the limit applies to the 179 deduction itself, on the business-use basis, and is not prorated. Take an $80,000 SUV at 90% business use. The basis is $72,000, 179 takes $31,300, and 20% MACRS on the remaining $40,700 adds $8,140, so year one is $39,440. The tool showed $36,936. The new test pins both $39,440 and the under-the-cap case ($40,000 at 60%: all $24,000 expensed).
Freelancer Schedule C categorizer: Uber Eats was deducted at 100% as travel. The travel rule matched uber before the meals rule could see uber eats, so delivery meals went on line 24a at full value instead of line 24b at 50%. While writing the test, a second form turned up that the audit had not named: card descriptors like UBER *EATS HELP.UBER.COM contain uber.com, which the travel rule also took. Delivery apps (Uber Eats, DoorDash, Grubhub, Postmates) are now checked before travel. An UBER TRIP line still goes to travel, and the test checks both.
Cache-Control checker: a safe header failed the CI gate. Cache-Control: public, no-store sent with Set-Cookie was reported as DANGER, a possible cookie leak, and exited 1. Under RFC 9111 §5.2.2.5, no-store forbids every cache from storing the response, so no cache can replay the cookie to another visitor. The finding no longer fires when no-store is present. The old test that expected this false alarm now checks a header that really is dangerous (public, max-age=60).
PWA installability checker: it claimed to agree with Lighthouse, and it doesn't. The README, the code header and the FAQ said this tool and Lighthouse "should never disagree". Lighthouse 12 removed its PWA category in 2024, and Chrome's published criteria (web.dev/articles/install-criteria) list a 512x512 icon and a start_url, which this tool treats only as notices. That claim is gone. Before this fix, a page missing both passed with "the same criteria Lighthouse checks". It now passes with "No hard install blocker found", followed by the two items the published list still asks for. One judgement call: the auditor wanted those two promoted to failures. I left them as notices, because Chrome's installer has accepted manifests without them and a CI gate that fails on an installable page is a bug too. Instead, the pass message now says exactly what is missing. A later audit may disagree.
World clock: already fixed. The Reykjavik midsummer bug (sunset after midnight, daylight shown as "-3h -51m") was fixed in PR #14327, with a test. That PR's audit never listed it in fix_prs. It does now.
What this says about the process
Three of the four wrong answers were domain rules: a tax limit, a tax line, an HTTP caching rule. A unit test written by the same author would not have caught them, because the author believed the wrong rule. Each one was caught by checking against a source the author did not write (Pub 946, the Schedule C instructions, RFC 9111). That is the same pattern as last week's reflection, where domain-rule and untested-core dominate the open defects. The fifth shows a smaller problem in the process: a fix that does not update fix_prs stays in the queue, and someone has to find it again.
37 high-severity defects are still open (python3 platform/ci/quality_audit.py queue --severity high). Most of the rest are research-design issues and untested core logic. They need more than a one-line fix.
These are free, MIT-licensed estimators, not tax or legal advice; check your own numbers with a qualified professional.