Public claims

No receipt, no claim. In public copy every figure, credential, promise and command has a receipt. A claim without one is removed, never softened into "about", "up to" or a vaguer label that implies the same thing.

A receipt identifies what supports the claim and how it was checked:

• a source with a locator: path:line, a document and section, the issuer's own listing;
• a command that re-derives the value on a pinned commit;
• the author's own words, with date and the exact text, supporting an attributed report rather than independent reproducibility.

Distinguish current reproducible results from historical or private-source reports. For a historical result, retain the original revision, date and evidence; a recount today measures today's state. If the source is private, say so in the public claim and identify its type and date without exposing the private link or material. A row can preserve the exact authorized author report, for example, "The author reported 42 collected cases on <date>; the underlying run is private and was not independently reproduced." Do not turn that report into a verified current test count. If disclosure is not authorized or adequate evidence is unavailable, omit the claim.

This skill never posts anything. Publishing is a separate decision for a human.

Procedure

1. Pin the reviewed revision and write its commit at the top of the fact sheet. Current recounts run on it; historical claims name their own revision and date.
2. Keep one fact sheet, FACTS.md, beside the drafts. A row holds the claim, its value, receipt and verification limits. Drafts quote the sheet and never introduce a new fact.
3. Find the receipt for each claim in every draft. No receipt: delete the sentence and log it in the changelog.
4. Check each figure against its stated basis. Recount current figures on the pinned revision; retain and verify the source for attributed historical reports. For pytest cases, run pytest --collect-only -q tests in a clean checkout of the pinned commit and record its exit status and collection summary, with configuration and environment. Counting def test_ lines counts matching definitions, not collected cases: parametrization, methods and collection rules can change the total. Collection also does not prove that the cases passed.
5. Check every universal. "Every", "never", "all" and "only" each need a receipt for the universal. When the claim is narrower, write the narrower sentence. Precise beats grand.
6. Show only real output. Quote the actual message text the tool prints; never write an imagined terminal session.
7. Tag unshipped work [PENDING] in drafts and keep it off public pages.
8. Use tokens for undecided names: {{NAME}}, {{REPO_URL}}, so one replace lands the final choice. A private link never appears.
9. Check length with tokens filled at their final length. Verify the platform's current counting rules and limits for the account and post type, including link handling; use its composer or documented counter.
10. Keep a changelog at the bottom of the fact sheet: what was removed, recounted or tagged, why, and on which revision.

Fact sheet row

Done when

Every claim in every draft has a fact-sheet row that supports its actual scope: current results re-derive on the pinned revision, and historical or private reports carry clear attribution and verification limits. Unreceipted claims are gone, not hedged; every [PENDING] tag has been re-checked; and the changelog names the revision reviewed.