Work
Machine-checkable verification first — quotes, figures, attributions, links, dates, units and denominators — with reproducible evidence. Then judgment-class critique, labeled as judgment: a standing adversarial battery, fairness-in-context, and a steelmanned counter-reading, on at least two independent frontier models.
Required trace
Every check emits an inspectable record of what was checked, how, and what could have changed the claim. “Ran, passed” never appears. Before the audit page publishes, it is mechanically checked against its own raw returns — every objection traced to a source return, every count recounted.
Constraint
Kills, changes, and blind spots — never verdicts. No survival counts, badges, or validation language. Per-claim calibration (claim type, evidentiary confidence, what would change the conclusion) is required reporting, not a verdict. Every page carries its execution-mode line: models, transport, independence discount.
Decision right
Audit pages and the tracked-claim ledger publish under the audit’s own authority. Three things always escalate to the author: any edit to published essay text, any claim that would bind the author’s personal conduct or identity, and anything the author has flagged.
Failure route
A floor-class defect found in a published essay stages a correction under the standing corrections regime and notifies the author. A wrong staged correction still costs the author time, so the audit’s false-positive rate is tracked on its own page.
Remaining risk
The audit’s zero-hit results inherit its refute-by-default bias, and its largest gaps are named on each page — including, on Issue 23, a fully specified disconfirming experiment that was not run. Timing is a disclosed trade: judgment-class defects the old pre-publication gates would have blocked now ship first and return as corrections.