Written for the analyst producing the work, and usable verbatim by the PM, the risk officer, or the allocator reading it.
1
Timestamp
Was the conclusion sealed before the outcome, somewhere you do not control? A git commit hash, a third-party email, an archived page. Work dated after the fact is a reconstruction, and should say so in the first paragraph, not the footnotes.
2
Provenance
Does every factual claim carry a source a reader can reach: publication, date, and a working link? “The bibliography is in the workflow” means the reader cannot check, and a citation the reader cannot check is a claim on your credibility, not the source’s.
3
Verification rate
What share of the model’s citations did a human independently confirm? Report the number. If no one checked, the honest rate is “unknown”, and that word belongs in the document.
4
Baseline
What would the dumbest defensible alternative have concluded? Consensus, persistence, or the same question put to the same model in one plain prompt. Edge is the distance from that baseline, and it cannot be computed if the baseline was never run.
5
Error budget
Before running, what result would have counted as failure? If nothing would, the exercise was a demonstration, and demonstrations belong in marketing.
6
Reproducibility
Model and version, data sources, run date, cost. A result that cannot be re-run cannot be audited, and a process that cannot be audited cannot be diligenced.
7
Disconfirmation
Did any pass of the work try to kill the conclusion? Agents that only gather support will find it.
8
Disclosure
Is every ex-post element labeled as ex-post, at the point of use? Partial disclosure in a limitations section does not license a headline that reads as live.