Keep an accurate history.

What actually happens to a record

Every record goes through the same seven steps. Nothing is skipped for convenience, and no step depends on someone deciding a statement deserves to be believed.

1. You paste a link. Anything said publicly — a post, an essay, a transcript, a filing.
2. We read the source from the platform itself. X and TikTok hand back a post’s own text through their public endpoints; articles and blogs are fetched directly. That response is stored word-for-word and fingerprinted; each record publishes a canonical extract of it and the fingerprint of the whole.
3. Software reads the statement and proposes the fields: who said it, the exact wording, whether it’s a prediction, a commitment or an opinion, and — if it can be settled later — what would settle it and by when.
4. The quote is checked against the source. Not for plausibility: we check the words literally appear in what the platform returned. If they don’t, the record is refused. A made-up quote attached to a real link cannot be published here.
5. The account is checked too. The platform tells us which account published something, and we read it from them rather than from whoever submitted it. You cannot attach one person’s post to another person’s name.
6. A person sees exactly what will be locked, and their only choices are lock or refuse. Review is not editorial power: the wording must match the source or the record is refused. Once locked, the anchored proof cannot be rewritten and the published record is append-only — corrections are appended in the open, by anyone’s challenge, never edited in silently. We also ask the Internet Archive for an independent copy — that often fails, and the record says so plainly rather than implying corroboration it doesn’t have.
7. It’s fingerprinted onto a public blockchain within seconds. Anyone can confirm a record hasn’t been altered since — the check runs in your browser and reads from a public node, never from us.

Afterwards the record stays open to challenge. Anyone can dispute it, and the person it’s about gets priority; we answer within 72 hours and resolve within 14 days. Corrections are added on top and remain published beside the original — the original is never silently edited, including when we were the ones who got it wrong. Where law or our published policies require a readable record to be withdrawn from display, the withdrawal is documented and the anchored proof remains.

What a record looks like

Every record carries its exact wording, its original source, when it was said and when we captured it, how the quote was checked, and a proof you can verify yourself against a public blockchain.

One registry, stated precisely

Not every record is the same kind of thing, and blurring them would overstate what some of them prove. The distinctions, plainly:

• Self-registered predictions — anyone may lock their own call here. What is proven is the timestamp and the words; the account’s identity level says how much the attribution is worth.
• Third-party records — limited to public figures and organizations speaking publicly, verified against the source itself.
• Pre-outcome records — registered while the answer was unknowable. Only these carry the locking guarantee, and only these count in track records.
• Historical records — a small, clearly labelled set registered after their outcomes were known, preserved for the source record only and excluded from every scored surface.
• Automatic vs evidence-reviewed outcomes — a locked test settles from a supported data source where one exists, or by a documented, named review applying the locked criteria where it doesn’t.
• Public proofs vs licensed data — every record and its proof is publicly checkable; the bulk corpus, entity graph and resolution datasets are licensed.

Why this matters

The record of what was said has never been easier to lose. Posts are edited or quietly deleted, links rot, video gets re-cut. And as synthetic media improves, “I never said that” becomes easier to claim and harder to disprove — for the people who said something, and for the people who didn’t.

Meanwhile more rests on that record than ever, with AI systems now answering questions about what someone said by reassembling whatever pages happen to still exist. When the record becomes unreliable, accountability doesn’t collapse dramatically — it quietly stops working. Nobody has to lie; it is enough that nothing can be checked.

How we do it

Four commitments govern every record. They are constraints on us, not features for you.

What makes this different

• No honesty score. Counts, sources and outcomes, each with the underlying records one tap away — never a verdict on a person.
• Accuracy is checked, not moderated. Nobody here decides whether a statement deserves to be recorded: if the words aren’t in the source, the record doesn’t exist. What Thonest’s software does contribute — the classification and the proposed test — is shown in plain language and confirmed by the submitter before anything locks.
• No token, and no chain of our own. We use a public ledger the way a notary uses a stamp: to prove nothing changed.
• Not for sale. Access to the data can be sold; influence over the record cannot. No payment removes a record or alters a resolution — and the archive is built so we couldn’t do it quietly even if we wanted to.

What we publish

• The record — 126 sourced statements and counting.
• Our methodology — how records are captured and resolved, and the limits of each layer.
• Cryptographic proofs — per-record proofs and every anchored batch in full, so the archive can be audited for retroactive edits.
• Public verification endpoints — individual records, per-record proofs and a completeness manifest, available without an account. Bulk corpus access, enrichment, entity data and resolution data are licensed separately.