Fewer false positives, defensible alerts on every disposition.

We start by inventorying current rules and pairing each with its false-positive rate, true-positive rate, and last-tuning date. Tuning happens scenario by scenario with a written hypothesis, a baseline run, and a comparison run; coverage loss is calculated against a held-out sample of historical SAR-confirmed cases. No rule is retired without a signed-off memo.

Every triage model ships with a model card, a scenario validation report, an alert-disposition audit log, and a written tuning policy. The model never auto-closes alerts; it ranks and routes them with a confidence band and a reason-code stack your analyst sees on screen. The examiner reads the same logs your QC team does.

When an analyst opens an alert, the system surfaces a dossier: customer history, related parties, transaction patterns, sanctions screening results, and links to prior alerts on the same customer. The analyst writes the narrative; the system assembles the supporting facts. Median SAR prep time drops, decision quality goes up.

Screening output flows into a review queue with confidence bands and a name-matching audit trail. High-confidence matches escalate; ambiguous matches route to a reviewer with the matching factors highlighted. Every disposition logs the reviewer, the timestamp, and the reasoning so a regulator can walk a sample end to end.

A binder, not a scramble. Architecture-of-record, scenario inventory, tuning logs, model validation memos, alert-disposition samples, SAR sampling QC, vendor risk packages, and the board-approved BSA / AML governance policy. We co-author the response to examiner questions on request.