How we verify

BSKiller is a nightly fact-check of the AI claims everyone is already seeing. Every claim gets one of four rulings with the receipts attached, and every claimant gets a permanent track record. This page is the full method, because a methodology you can't audit is just branding.

THE RECORD SO FAR: 291 claims · 2468 receipts · nothing unsourced

The four rulings

Every checked claim ends in exactly one of these. No composite scores, no vibes:

• BSnot supported by its own sources
• CONTESTEDcredible experts genuinely disagree
• TRUE, BUTliterally true, materially misleading
• VERIFIEDwe tried to break it and could not

What we check (the coverage law)

We fact-check the claims people are actually seeing, not the claims we happen to find. A trend desk reads the biggest AI amplifier newsletters and the loudest forums every day; any claim echoed by two or more independent amplifiers becomes a mandatory candidate for that night's issue: we stamp it, or the issue notes say why not. A fact-checker that misses the claim everyone saw is not a fact-checker. Those newsletters and forums are discovery sources: they tell us what to check, never what is true.

Evidence and discovery: two tiers of source

We keep a lead and a proof apart. Every source we use is in one of two tiers.

• Evidence. Every source that is not on the discovery list: primary documents (filings, papers, model cards, official statements, datasets, regulator and court records) and reporting by named outlets. A ruling is built from this tier. The test tells the tiers apart by web address, not by reading the page, so a person still judges whether a source is primary or original.
• Discovery only. The listed hosts: Hacker News, Reddit, GDELT and the three AI newsletters our trend desk reads (The Rundown AI, Ben's Bites, TLDR AI). These show what people are seeing and arguing about, which is how we find claims worth checking. They are opinion and attention, not evidence. A discovery link can sit beside a ruling to show what readers were saying, but it is never the evidence a ruling stands on: a ruling with only discovery links fails the release.

A test enforces the floor rather than habit: the release fails if a ruling has no evidence-tier source, that is, if its sources are discovery links alone. Everything above that floor is editorial: the receipts described below, and a person judging whether each source is primary and independent.

How a claim becomes a ruling

1. Extracted, not summarized. We quote the assertion the claimant actually made, attributed and dated, with a link to where they made it.
2. Receipts, plural, including a hostile one. Independent sources across separate domains, and at least one chosen to argue against our ruling. Every source URL is checked live the day we publish and archived so a receipt cannot quietly rot. Every quoted line is matched to its source word for word. Rulings without receipts don't ship.
3. The named mechanism. Where a claim misleads, we name HOW: the specific trick (reference-class swap, denominator hiding, self-marked benchmark, and so on), because "misleading" without a mechanism is just an opinion with a scowl.
4. An adversarial gate, then a second, independent one. Every story passes a red-team pass and a separate automated review by two independent reviewers before it can ship. A factual flag from a single reviewer is re-run to confirm it isn't noise; a story that keeps failing is quarantined: publicly absent, never quietly published.
5. Falsifiable, with a stated flip condition. Where a ruling depends on something that could change, the claim page says what evidence would flip it, and Promise Watch tracks the claims with deadlines until they resolve.

The six questions about a company

The four rulings above rate single claims. A company dossier rates something else: how the company behaves, on the same six questions for every player. Each question is rated solid, mixed or weak, or "not enough evidence", from dated receipts, and a rating with no receipt is not allowed. A page that shows a count of these ratings says so; it is a count, not a score.

1. Launch claims vs independent tests. When they say a model is better, faster or cheaper, do independent evaluations agree?
2. Money claims vs the filing. Do their valuation, revenue and user numbers survive contact with primary disclosures?
3. Promises kept. Did the things they announced with a date actually ship, on time, as described?
4. Safety and incidents. When something went wrong, did they disclose it quickly and honestly?
5. Transparency. Do they publish what an outsider needs to check them?
6. Legal and regulatory record. What do courts and regulators say about them? A lawsuit being filed is an allegation, not a contradiction: pending suits alone rate Mixed at most. Weak needs a ruling, settlement, fine or regulator finding against them.

Track records: the memory

Rulings accumulate. Every claimant (company, executive, pundit) has adossier: their claims over time and how each ruling landed. That is the instrument: before you trust a word, check the speaker's record. These records are free to read, with no lookup limit.Membership opens full checked stories and the written archive. Two designated sample stories are public; signing in opens the edition’s designated free story. Source links and corrections remain public.

Corrections are loud

Fixed claims keep their history; we never silently edit a ruling. If you can show a ruling wrong, it changes, and the change is public: email[email protected]with the claim ID and your evidence. If you're right, the correction ships with credit.

Built to be quoted

Every ruling ships twice: once as the page you're reading, and once as machine-readable ClaimReviewdata embedded in the page, the same structured format Google's fact-check results use, so a search engine or an AI assistant can quote our rulingwith the rating, the date, and attribution back to us, instead of guessing. If a machine repeats one of our calls, it repeats the receipts too.