Accuracy, in the open
Every verification PaperTrail emits must ground each flagged span in an exact, verbatim substring of the cached source text — an ungrounded span is a hard failure of the harness, not a soft miss. This is a stronger guarantee than a confidence score alone: it means the tool cannot claim the source says something the source does not literally say.
No accuracy numbers have been recorded yet.
Accuracy is measured against pinned-PMID labeled fixtures in tests/fixtures/demo-claims.json — each fixture pins a real source by PubMed ID, an expected discrepancy type, and the exact substrings the verifier should flag. The harness scores the full extraction → verification → reconciliation pipeline deterministically against those labels.
Numbers appear here after running npm run eval against a live database with the demo sources ingested (npm run ingest:test-set). Until then this page shows no metric rather than an invented one.
On the span-grounding guarantee
PaperTrail enforces that every flagged span maps back to a verbatim substring of the cached source, so its span-grounding rate is 100% by construction rather than an estimate. For context, published evaluations of general-purpose retrieval and citation tools have reported quote/attribution match rates well below 100% (for example, studies of AI answer engines have found substantial fractions of citations that do not support the associated statement). That figure is cited only as external context on the problem space — it is not a measurement of PaperTrail. PaperTrail's own numbers are the ones shown above, and only after npm run eval has been run.