Proof, not testimonials.

Our own site, measured the way yours would be

Before this repositioning, this site carried a 2 MB autoplaying video and a 1.14 MB poster image in the critical path, eight pages each with their own drifted copy of the design tokens, and a bright cyan used as body text on white in about thirty places at roughly 2.1:1. All of that was found by tooling rather than by eye.

This site is swept with the same tooling on the same schedule as a client engagement. The current result is zero failing elements across all sixteen public pages, from one axe-core 4.13 run at 1280px taken immediately before publishing. The bright brand cyan is kept where it is correct, on the navy nav bar at 8.4:1, and replaced anywhere the computed ratio said otherwise.

That run also left 159 checks undecided. Automated tools decline to judge color contrast when they cannot establish what sits behind the text, and on a site with a photographic hero and gradient bands that covers a great deal of the page. Undecided is not a pass, so every one is settled by hand: the text is hidden, the area behind it is captured, and the ratio is computed against the worst pixel in the box rather than the average. All 159 pass. That second pass is the part most audits never run, and it is included in yours.

Every finding carries its own proof, enforced in code

The engine behind the free site check is built so a finding cannot exist without the nine fields that make it actionable. The record builder throws rather than emit an incomplete finding, and a test asserts that for each field in turn. Severity definitions live in the same module the website publishes, so the page and the code cannot drift apart.

Two integrity rules are enforced rather than intended. A check that could not run is reported as not measured, never as a pass, and never folded into a coverage score. And a target’s own robots.txt decides whether we fetch it at all, with a robots.txt we cannot retrieve treated as unknown rather than as permission.

Earlier work applied the same idea to a lead-scoring engine, where a scoring rubric and a written QA plan were maintained alongside the code so any score could be traced to the evidence that produced it.

64 tests, and two bugs found by asking where a number came from

We run an internal trading system. It is the hardest test of QA discipline we have, because a silent defect costs money rather than goodwill, and it produced two findings worth describing.

Two writers for one number. Account equity was being set by two different code paths: a broker heartbeat and the system’s own estimate. Because position size, risk caps and the circuit breaker are all percentages of equity, a drifting estimate would mis-size every trade. The fix was a rule rather than a patch: any number that drives sizing has exactly one authoritative source, named in code. It was found by asking where a number came from, which is the same question an audit asks.

A confident score from a failed measurement. A performance check was recording a 79-second timeout as a near-zero speed score for a site that answered in 0.36 seconds. Left alone, that would have told a business their website was slow when it was not. It is the origin of the rule that a timeout is never a zero, now enforced in the audit engine with a test named after it.

Evidence-gated features. A position-sizing multiplier stays capped at 2x, whatever its configured ceiling says, until 25 closed trades have actually tested it. The setting does not get trusted before the log can back it up.