
Markets get rigged in ways that keep changing. A trader posts orders they never intend to fill to fake demand (spoofing), stacks fake depth across price levels (layering), trades with themselves to invent volume (wash trading), or pushes the closing price to mark their own book. Regulators write a rule for each one, but the rules sit still while the tricks keep moving. A small tweak to a known trick can walk straight past a rule written for last year's version of it. Alpha & Oversight runs two desks that never share a model or a memory and only talk through Band. On the R&D desk, a frontier adversary invents the next evasion. On the Surveillance desk, seven agents work the case: anomaly detection, a specialist pulled in for the job, a prosecution-versus-defense debate, and an adjudicator who settles the contested inputs. The agents argue. They never decide the verdict. A plain deterministic rule engine does that, the same way every time, and cites the exact rule and metric that tripped. The part we care about most is what happens when the adversary wins. A novel evasion slips the four seed rules, so the engine passes it, but the case still looks wrong, so it escalates to a human instead of closing. The human confirms it's real, and that one click derives a new rule from the case, replays it through a regression gate to prove it now flags, and codifies it. Active rules go from four to five and the case flips to FLAGGED. Every handoff and verdict is hash-chained into an audit ledger you can replay, with each entry tied to a real Band message. The four seats on the adversarial boundary run four different model families, so one model's blind spot doesn't quietly become everyone's
19 Jun 2026