How we verify

BSKiller runs every book and paper through the Zero Hallucination Protocol. The rules, in full, because a methodology you can't audit is just branding:

The rules

1. Every claim is extracted, not summarized. We quote the assertion the author actually makes, at most 40 words, always attributed.
2. Every verdict has receipts, and more than one. At least three independent sources across three separate domains, and at least one of them is chosen to argue against our verdict, not for it. Every source URL is checked live (a real HTTP 200) the day we publish and archived to the Wayback Machine, so a receipt cannot quietly rot. Every quoted line is matched to the source word for word before it ships. Verdicts without receipts don't ship.
3. Verdicts are falsifiable. holds / failed / contested / unsupported / needs context, each with a stated confidence. If you can show a verdict wrong, it changes, and the change is public.
4. Corrections are loud. Fixed claims keep their history. We never silently edit a verdict.
5. Criticism, not substitution. For in-copyright works we publish commentary under fair dealing (criticism & review): quote budgets are hard-capped (≤40 words per quote, ≤150 per chapter, ≤1% of the work) and machine-enforced. Our editions are the argument about the book, never a replacement for it.
6. Public domain is restored, not exploited. Restored Editions carry a public rights dossier proving provenance.

Current corpus: 632 claims, every one at a stable URL with its audit status shown.

Audit status: what "verified" means here

Every claim shows one of two statuses. ✓ Verified: passed the full protocol: quote grep-checks against source, receipt validation, and an adversarial red-team pass in an independent session. ⚠ Extracted, re-audit pending: produced by our extraction pipeline with receipts attached, but not yet through final verification. We publish the distinction because hiding it would be exactly the kind of BS we exist to kill. Unreviewed claims are re-audited before any work is promoted, and the count on the homepage tracks progress.

The BS Index, defined

The number on the front page is not a mood. It is a weighted average you can recompute yourself. Every AI claim we check is scored on a 0-to-100 scale for how much bullshit it carries:

• Verified = 0. Survived the receipts intact.
• True, but = 40. Technically true, sold with the load-bearing context left out.
• Contested = 65. Credible evidence on both sides, including the claimant's own record contradicting it.
• BS = 100. Did not survive contact with the evidence.

The BS Index is the average of those scores across every claim in the current issue. Verified-heavy issues pull it toward 0; an issue full of half-truths and puffery pushes it up. The verdict mix behind today's number is printed right under the gauge, so the arithmetic is always in view. Change one verdict and the index moves by a predictable amount. That is the whole point: a bullshit meter that is itself impossible to fake.

Built to be quoted

Every verdict ships twice: once as the page you're reading, and once as machine-readable ClaimReviewdata embedded in the page. That is the same structured format Google's fact-check results use, so a search engine or an AI assistant can quote our verdict on a claimwith the rating, the date, and attribution back to us, instead of guessing. We want to be cited. So we made the verdict impossible to misread: the claim, the rating from 1 to 5, the sources, and who said what are all in the markup. If a machine repeats one of our calls, it repeats the receipts too.

Found an error?

Email corrections@bskiller.com with the claim ID and your evidence. If you're right, the correction ships with credit.