Published accuracy.

These numbers are from the run on 2026-07-27, and that run is behind the code.

The suite calls the live model, so it is a deliberate manual run rather than something that fires on every commit. Guards have landed since: retrieval before judgement, the confidence rubric shipped in the prompt, the retraction filter, and quote verification with a full-text fallback. We expect the false-attribution rate to be materially lower on the next run, and we are not going to write a better number here before we have measured it. The dataset is also 20 hand-labeled cases, which is small enough that one case moves a rate by five points.

Of the cases we labeled as having no real supporting source, how often the engine returned one anyway AND rated it STRONG or DECENT. This is the number that matters, and the only one where lower is better.

How it is measured
Counted over the labeled no-source cases only. A source returned with a low verdict does not count against us; asserting confident support that is not there does.

When a real supporting source exists in the open literature, how often we find it.

How it is measured
Counted over the cases labeled as having a findable source. Bounded by what open academic databases actually index, so it can never reach 100% on the whole world.

How often the verdict lands in the set a human labeled as acceptable.

How it is measured
Each case carries a SET of acceptable verdicts, not one right answer, because STRONG and DECENT are often both defensible on the same paper. That makes this a lenient measure: a miss means the engine landed outside every answer we would have accepted.

Of the verbatim quotes returned, how many are genuinely verbatim in the source.

How it is measured
Verified by normalized substring match against the retrieved text, with a full-text search as fallback. A quote that fails is dropped and the citation is kept, so this measures the verifier, not the writer.

Honest caveats: this is a labeled suite, not the whole world. Recall is bounded by what the open literature covers, the dataset is small and hand-built, and a good score is not a guarantee on any single claim. The bar is "good enough to win an argument or pass an undergrad citation," not a flawless systematic review.

Receipts are alive; citations are not

A citation points at a verified fact: a real paper said this, on this page, in this year. A receipt is a different animal. It shows evidence-backed argumentation, the best current reading of both sides, and it is built to change. New sourcing can strengthen it, a better counter-argument can soften its verdict, a retraction can downgrade it overnight, and a shift in the evidence field can flip what CONTESTED means for it. Every change lands on the receipt's public, append-only history, so you can always see what it said before and why it moved. A receipt that never changed would not be more trustworthy. It would just be less watched.