audit-post

Audit-Post — the ONE correct, unskippable procedure for auditing a published post

What this is

The end-to-end procedure for auditing one already-published post to a scientific-organization standard. It chains the adversarial skills and — critically — re-runs the auditor on the CORRECTED post to confirm it now passes clean before committing. We are a scientific organization; we do not ship missteps. Run EVERY step, in order, for EVERY post. No light/inline version. Fewer posts done FULLY > more done shallow.

Defined 2026-06-29 after the owner caught two failures: (1) shortening the audit to a single citation verifier + inline reasoning (missed a "Bayesian" mislabel, a strawman baseline, a wrong-estimand, and a "textbook dressed as our law" overclaim across posts #7–#10); (2) dropping the re-audit-after-fix step — I fixed and committed WITHOUT re-running the auditor on the corrected post to confirm the fix is clean.

Updated 2026-07-01 after a THIRD failure: audits #16–#19 all ran stress-claim + verify-claims but never ran storm-research at all, because this file — written 2026-06-29 — predates the owner's 2026-07-01 PERMANENT standing gate ("VALIDATE → STORM → AUDIT → VERIFY, storm is dominant, never skip, applies to everything not just flagship posts"). The gate overrides every skill file by design, but relying on remembering an overriding rule from a different document is exactly the failure mode that gate exists to prevent. STORM is now Step 3 below, hard-coded into the numbered procedure itself — not something to recall from memory. If a future owner rule changes the required order again, update this file itself, don't leave the gap for the model to bridge silently.

Updated again 2026-07-01, SAME DAY after a FOURTH failure, caught within an hour of the third: the "run the /seo Mode-A check" instruction was a clause buried inside a bigger step (old step 6, mixed in with bilingual-fix instructions), and in practice it was never actually executed as a distinct checklist pass on audit #19 — I made ad-hoc meta/JSON-LD/internal-link edits without running the LLM-summary self-test, without checking the Three Kings against the char-length rules, and without requesting re-indexing. Writing a step into a paragraph is not the same as writing it as a step. SEO validation is now its own numbered step (7), with the specific /seo Mode-A checklist items spelled out inline so there's no clause to skim past. Pattern to notice going forward: if a "run X" instruction is a sub-clause of a different step instead of its own step, that's the shape of thing that gets silently skipped — split it out.