The method: how the receipt gets made.
Search first, judge second. The engine never decides whether a claim is
true from memory; it retrieves real papers and reasons over what it actually found.
The pipeline, in plain English
- Your claim becomes several search queries, aimed both FOR and AGAINST it.
- We retrieve real papers from open academic databases (with citation counts, and
retracted papers thrown out).
- The model sorts what was found into supporting and contradicting evidence and
calls the balance: SUPPORTED, CONTESTED, REFUTED, or INSUFFICIENT EVIDENCE.
- A deterministic score (not the model's opinion) computes confidence from source
agreement, citation strength, replication, and counter-evidence.
- Quotes are only shown when they exist verbatim in the retrieved text. A missing
quote never blocks a citation; a fabricated one is the single unforgivable failure.
The receipts correct themselves
Every receipt keeps an append-only history. If a cited paper is later retracted, a
nightly scan catches it, downgrades the verdict, and logs the correction publicly.
Anyone can challenge any receipt, but only with a source; opinions don't move verdicts,
papers do. 0
claims are on the record so far.
Published accuracy →
We run a labeled evaluation suite (false-attribution rate, recall, quote fidelity,
verdict accuracy) before every release, and we publish the numbers here; including
the imperfect ones. Nobody else in this category shows you their accuracy. We think
that's exactly the problem.