What an audit of this site found, and what I fixed

Links

75 of 697 were dead or pointed somewhere other than where they claimed; I repaired 59 and removed 22. The failure had a shape: 32% of ISBN-path URLs and 55% of Penguin Random House numeric IDs resolved to an unrelated book. Gendlin's Focusing went to a book called The User Illusion. Every DOI, PubMed link, and government or guideline link was correct.

Evidence claims

243 checked, 52 corrected. Multisystemic Therapy's effect size was inflated by more than double. Solution-Focused Brief Therapy carried an effect size its cited source cannot produce; that paper computes none. CBASP was listed as NICE-recommended for chronic depression; NG222 does not mention it. Positive Psychotherapy cited "WHO"; the source was the founder's own trade association. Case series and single-arm pilots were called randomized controlled trials. None of these look like errors. They look like scholarship, which on a teaching site is worse.

Prose

256 changes across 132 modality entries. Somatic Experiencing said Peter Levine studied with Wilhelm Reich, who died when Levine was fifteen. ACT said Hayes never embraced "third wave," a term he coined. NARM's ontology had been pasted into the Wilhelm Reich entry. Sixteen entries credited one founder where the data named two, and the dropped name was usually a woman's.

Vignettes

The worst-affected content, and the site's most distinctive feature: 161 changes across 121 of 156 formulations. FBT asserted the family causation the model exists to reject. Motivational Interviewing was built on a model its founders disavow. A client with DID was given the wrong level of structural dissociation. Students read these as a modality's real reasoning, so an error teaches the wrong model under the right heading.

Debates

All 19 changed, 84 fields. Two quotations were fabricated, including "between stimulus and response there is a space," which is not Frankl's. Judith Herman was cast as defender of the diagnostic system she spent a career criticizing.

Conditions, questions, glossary

83 changes. Six prevalence figures were materially wrong. Three entries had clinical safety gaps: the eating disorder entry omitted medical co-management and refeeding, the perinatal entry omitted postpartum psychosis, the couples entry omitted screening for intimate partner violence.

And the glossary defined epoché. It credited Husserl with a term he borrowed from Pyrrhonian skepticism, and gave the loose qualitative-research sense rather than the suspension of the natural attitude, which it never mentioned. I had not reread the entry since it was generated.

What held

All 103 DOIs resolve to the correct paper, as does every PubMed, government, and guideline link. The controversy sections, which document ethical allegations against named practitioners and are the riskiest content here, came back 85 of 98 supported, with no legally sensitive findings and no unsupported claim about any named person, several corroborated against court records and licensing board files. The most recent entries, written slowly and checked as they went, showed no confirmed errors.

How this happened

I generated content in large batches with AI assistance and published it without checking the citations. The tool did what these tools do: it wrote fluent, plausible references. Verifying them was my job and I did not do it. Every page's footer said the content had been reviewed for clinical and scholarly accuracy. It had not been. Where the content used identifiers a model cannot invent, it held. Where it used strings a model can generate, it did not.

What changed

Everything above is fixed in the data, not queued. Where no correct link could be verified, the URL came out and the citation text stayed: a citation without a link is ordinary scholarship, a link to the wrong book is a false claim. Nothing was guessed; inventing a fix would repeat the original failure. The site now prefers DOIs and other resolver-backed identifiers over publisher URLs.

What is still open

Nine citations remain unverified behind paywalls; the sources are real, but unchecked is not clean. Three structural items wait on my judgement: twelve contraindication fields listing therapist training gaps and clinician misuse rather than contraindications, five modalities filed in a lineage they do not descend from, and two grief entries that are the same protocol entered twice under its old and new names.