EVIDENCE / REPRODUCIBLE & SYNTHETIC

Show the score.
Keep the failure.

This page runs the actual pinned local model. Four calibration examples, twelve held-out cases and one adversarial case were written before the first inference run. Threshold 0.82 and runner-up margin 0.06 were fixed in advance. No cases are dropped when the model misses.

Labels are authored expectations. These short English fixtures are not a user study or an estimate of performance in real documents. A “repeat” candidate can still be wrong. The role-swap case deliberately probes a limitation of sentence similarity.

Ready. This will perform real CPU inference in your browser.

No run yet.

Engine-only tests separately exercise malformed vectors, input limits, quote preservation, ambiguity and match reuse. The app keeps every current line even when a similarity suggestion is wrong. See README for the exact verification commands and limitations.