What Two AI Reviews Found
One review against my written standards, one blind first-visit pass, and a cross-check that kept both honest. What each caught, what each missed, and what changed on this site because of it.
The review that knew my standards missed the two defects that mattered most. A blind second pass caught both within minutes of arriving the way a stranger would: 19 of the hub's 27 "see the ledger" links opened nothing, and across all 55 verified documents, not one named source was an actual link. A cross-check then reproduced every countable claim before anything got fixed, and the structural fixes shipped behind build gates that fail on a return of their defect class; what the gates cover and what they do not is stated below, in plain sight.
How the review ran
Three passes, one morning. First, one AI tool reviewed the site against my written content standards: the voice rules, the design language, the publication contracts. Second, a different AI tool reviewed the live site blind: the first review was withheld from it, and it arrived like a first-time visitor with the standards as its only map. Third, the second tool cross-reviewed the first, reproducing or correcting every claim it had made, number by number.
The boundary, stated plainly: this was an internal cross-model review, two AI tools working inside my own system under my direction. It was not an external audit, and I am not calling it one. The value is in the method and the shipped fixes, not a borrowed credential.
What each pass caught, and missed
Five generations of page templates, 55 of 60 inner pages with no visible author, a sitemap nine documents behind the site, and a public vocabulary split across internal workflow terms. All real, and all later reproduced by the cross-check.
The blind pass found what the standards pass walked past: the hub's most trust-bearing interaction, the ledger link on each card, was broken on 19 of the 27 cards that offered it, because an unvalidated flag guessed at anchors that were not there. And the evidence ledgers named their sources as inert text: zero external links across every verified document. A site whose promise is verifiable proof was asserting, not showing.
Of 22 claims in the standards review, 9 confirmed exactly as stated, 13 held in substance with corrected scope, numbers, or priority, and none overturned outright. The corrections were the kind that matter in the small: 28 cards carried an unexplained tally, not 29; one page title did carry the site's name; "outside every gate" was too broad because two structural checks already ran.
The pieces that most directly show what I do sat at grid positions 20 through 22 of a flat catalog. Independently, both reviews said the same thing: the proof of practice should lead, and the library should sit beneath it.
What changed
The reviews checked the public evidence interface: whether the proof is present, linked, consistent, and honestly labeled. They did not re-verify the truth of every underlying research claim, and no build gate ever will; claim truth stays human work, on a refresh cadence each page states.
Two smaller findings were open when this note first published and closed later the same day: the adoption charts now carry an open-full-size link beside each figure (a phone can inspect the on-chart source footer at real scale, and the page's own source register was always readable), and every inner page now loads the site's self-hosted fonts instead of a third-party font service. Both fixes are dated here rather than silently absorbed, because that is the point of this page.
What to steal from this
A reviewer who knows your standards checks your compliance. A reviewer who arrives cold checks your promise. The second kind found what mattered here, in minutes.
The cross-check worked because the reviews argued in numbers a second tool could reproduce: 19 of 27, 9 behind, 55 of 60. Adjectives cannot be cross-examined.
Most of the repairs above came with a build check that fails if that defect returns: ledger targets, routes and redirects, the page contract, the retired vocabulary in page text and shipped data alike. The one gap this page originally named is now closed too: the evidence retrofit landed the next day, so every Verified ledger links its sources and a presence gate fails the build if one goes bare. A fix without a gate is a snapshot; the gate is the fix, and an ungated fix should say so.