Scoreboard
Not yet measured. Run npm run core:eval and rebuild the site.
Method
Positives are questions generated from claims first valid in a recent window, with the expected answer taken from the latest stored claim. Negatives are fabricated entities, nonexistent NCT ids and years with no data; the correct behavior is to abstain. Wrong rate is reported among answered questions, alongside coverage, with 95% Wilson intervals.
Limits
- Positives test the pipeline (parse, resolve, retrieve, render, verify). They do not test whether the sources are true.
- Core answers only within ingested domains and attributes.
- Copy-only rendering prevents invented values, not wrong reasoning or aggregation.
- Any baseline shown is a local model on money questions only, graded within 0.5%.