Contract Check

How well the agent works

Measured on a test set of generated contracts with a known answer key, some of them honest. A finding counts as correct when the trap type matches and its quote overlaps the labelled fragment.

No baseline yetThe numbers appear here after the first evals run. Until then this page does not show sample figures, so nothing on it is made up.npm run evals -- --split testWrites evals/results/latest.json: precision, recall, quote accuracy, false alarms on honest contracts, cost and latency, per model and per trap type.