On 300 synthetic index queries, same-answer recall increased from 22/150 to 117/150 (14.7% to 78%), with two wrong answers among 300 queries at both selected operating points. This measures retrieval behavior at the selected thresholds. A separate 30-query live replay served seven requests without another model call; no latency improvement is claimed.

Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.