A passing check is a claim, not a fact. On 18–19 July 2026, an audit day on this site revealed four serious defects — a page giving away a paid book, a contact form losing every enquiry, a blank site map, and a coherence metric mostly measuring text length — all behind checks that had already passed. The checker was Claude, working as the site's engineer. The thesis: checking again re-runs the instrument; checking differently tests it.

Eight failure modes are catalogued. 1. Checks that cannot fail: a script printing "BAD" twelve times then concluding PASS because a failure flag died in a subshell. 2. Wrong dialect: checks written for source code but deployed runtime differs (e.g., `href="/courses"` vs `href="/courses/"`). 3. Looking where the light is: auditing source where links exist, but failure lives in rendered page. 4. Measures that measure something else: "coherence" score correlated with length, not structure. 5. Test sets authored by the hypothesis holder conceal confounds. 6. Experimental designs that erase the effect they hunt. 7. Silence read as success: failed submissions swallowed by catch blocks. 8. Documents drifting from data: a page claiming steel and pine benches sit 9 °C apart at skin, but the underlying table was edited without updating the prose.

The fix: write sentences as tests — one assertion per claim, named after the sentence it defends. When assertions fail, they force truer statements.