{% extends "base.html" %} {% block title %}Synthetic Generator{% endblock %} {% block content %}
Runs the real synthetic eval loop over the committed Fabricate battle-test corpus (550 labeled conversations) via the existing fabricate_adapter + compute_detection_rates. Optionally ingests events into the audit store.
Detection is measured only over conversations that declare each violation; false positives are measured on the disjoint compliant population. ATR-01 is excluded (untestable in transcript replay — verified via the failure-injection harness instead).
No run yet.