The Final Audit

One procedure closes the course — take any capability claim published this month and run the whole kit against it, in writing, to a three-clause verdict.

Step one: quote the claim verbatim — audits of paraphrases audit the paraphrase. Step two: expand it into the three clauses. What is the task, as measured — which metric, which target, which passing line? What distribution was tested, and how far from it will the claimed use sit? What conditions were attached — curation, retries, human assistance, stakes? Step three: walk the ten questions, by number, writing each answer or writing "unanswered" — an unanswered question is a finding, not a gap in your notes. Step four: if the claim involves autonomy, multiply along the chain — ask for the whole-chain number, and treat its absence as the flattering factor left unmultiplied. Step five: if the claim forecasts, hold it to a named mechanism — which wall must move, and what would evidence of the moving look like, observably, in behavior or benchmark custody rather than in tone? Step six: write the verdict as one sentence of the fixed form — demonstrated this task, as measured by this metric, on this distribution, under these conditions — and notice how the sentence's honesty is enforced by its grammar: every slot demands a fact, and slots left empty stay visible. The procedure is symmetric by construction. Run it on a vendor's boast and on a columnist's dismissal in the same sitting; whichever survives with more slots filled has earned the belief. Nothing in the six steps requires anything beyond this course — which was the course's promise.