Capital-markets surveillance · MAR / MAS

The order book decides. The model only writes it up.

A surveillance desk drowns in alerts at ~99.99% false positives. Here, deterministic code computes every manipulation signal, cancel-to-fill, layering depth, close-window concentration, news-to-trade latency, and assigns the disposition. The language model is downstream and read-only: it drafts the STOR strictly from flagged events, each clause cited to an order ID and timestamp, and writes "abstained, no evidence" where there is none. No report files without an analyst's approval, and every decision lands in a hash-chained, replayable ledger. ⚠ synthetic data, illustrative thresholds

20 / 20

manipulations caught

false alarms / 180 benign

100%

STOR clauses cited-or-abstained

23 / 23

ledger replay verified

Order book & comms

ACME · 2026-06-30 · synthetic

17:00:00 · drag to replay the session

Alert queue

rule-based disposition · code decides

STOR draft

select an alert

Select an alert in the queue. The engine drafts a Suspicious Transaction & Order Report, every clause hyperlinked to the event that justifies it, or marked abstained where no evidence exists.

Labelled suite

200 episodes · 20 manipulation + 180 benign-but-suspicious

A measurement, not a claim. The generator emits a held-out, ground-truth-labelled suite; the same deterministic rules score every episode. A third party re-runs the seed and gets identical verdicts. Run it to watch the confusion matrix fill.

AI Control Ledger, append-only, hash-chained empty, approve or reject a STOR

Each entry's hash covers its contents AND the previous hash (FNV-1a, computed in-browser). Edit any record after the fact and the chain breaks at exactly that row, that is what "tamper-evident" means.