Review ledger

How the review works

1. Every checkable statement is tagged in the text with its source: a cell of the evidence plane, a notary root, or an external document pinned by its SHA-256 fingerprint.
2. The reviewer re-reads every source. Each statement is verified, contradicted or unverifiable, with the observed value and its pointer. The figure the reader sees must match the evidence.
3. It also flags figures outside a tag, any investment advice (French, English, Arabic), and any difference between the French reference and its translations.
4. A contradiction, advice, a translation mismatch or an invalid contract blocks the text. Otherwise it is "ready for a human": only an authenticated reviewer approves it, for the exact fingerprint of the text, and the decision is published by role. The database recomputes the verdict and refuses an approval of a blocked text.

Verified share of declared claims

The number of claims the author declared that the evidence verified, over the number of distinct declared claims in this text (the same evidence counts once). A claim is verified only when its text is the controlled rendering of the verified typed assertion; any other wording is left to a person. It is a count, not a sample: no interval is given. It is not a probability that the text is true. Any figure outside a tag blocks the review.

A language model’s confidence in itself is excluded: it is poorly calibrated. A model pass, when it runs, can only add warnings; it verifies nothing and lifts no block. The meaning of a translation is left to a native-language reviewer.

Every review is sealed

The text’s fingerprint and the results’ fingerprint form a Merkle leaf; the day’s root is published in the public notary. Any later change to a reviewed text or its review is detectable.

Full ledger as JSON → · /api/v1/notary →