Before version 2.0.0 shipped, we stopped building and started prosecuting.
Four audit panels — logo, implementation, brand and documentation, design and themes — were convened with the same instruction: assume the previous team was wrong, verify everything yourself, and rank what you find. They rendered marks at real sizes, re-derived geometry from the generators, recomputed contrast ratios by hand, diffed the copy on the site against the deck, and read the emitted HTML rather than the source that claimed to produce it.
They found real things.
The titanium material treatment rendered a gray rectangle behind the mark — an unclipped noise field in an SVG filter, invisible in the code, obvious in the render. Fixed in the generator, not the artifact.
Two panels independently found the same blocking defect: with JavaScript disabled, every headline on the site sat at opacity zero, waiting for a scroll observer that would never run. The content was in the HTML — crawlers saw everything — but a human without JS saw a dark field. The fix is a one-line gate: hiding now applies only after a script has announced that scripts run. Progressive enhancement is a checkbox everyone claims; artifact inspection is how you find out.
The documentation panel caught the subtlest class of failure: truth decay. The performance page published measurements from a framework version we had upgraded past; a governance doc described a link-integrity gate as "planned" three days after the gate went live; a count said eleven where the suite said twenty-three. None of it was dishonest when written. All of it was false when read. Documentation that states measurements now cites the audit that made them.
And the bundle: first-load JavaScript came in at 198.3 kilobytes gzipped against a 200-kilobyte budget. We recorded the 1.7-kilobyte margin as a non-regression invariant instead of celebrating it. A budget nearly spent is the useful kind — every future addition to the first-load path must now displace something, which is exactly the argument a design system should force.
The audits are in the repository, verdicts and all, next to the fixes they caused. A brand that says "it shows its work" ships its cross-examination.