no ambiguous transactions yet — the sieve is clearing traffic
Offline evaluation (held-out test set)
This is a recorded run, replayed in the browser. Every number shown — sieve probabilities, VQC and QAE
scores, verdicts and Shapley attributions — was produced by the trained models on held-out transactions and
written out ahead of time by scripts/export_static_demo.py; the page replays those records
rather than scoring in real time, which is what lets it run on static hosting. The feed is also enriched to ~8% fraud
and ~22% ambiguous cases for visibility, against a real prevalence of 0.172% fraud and 0.6% ambiguous, so the
"cleared synchronously" figure here sits well below the 99.4% measured on the real distribution. To watch the models
score live, run the FastAPI server from the repository.