Most fraud systems ask institutions to trust their output. QUORUM is built so that each decision can be checked: deterministic rules decide in real time, AI models review afterwards under human control, and every decision is recorded in a form a third party can verify.
QUORUM's adversarial tooling generates red-team test cases and proposes new rules from what it finds. Those proposals never take effect on their own: they run in shadow mode, are reviewed by an operator, and are promoted only with WebAuthn step-up and, where required, a multi-party FROST approval.
Three locally run, fine-tuned models analyse cases from different perspectives — behaviour, financial context and evasion — and an arbitrator reconciles them when they disagree. They run asynchronously, so an AI outage never stops real-time decisions, and their output is evidence for analysts rather than an unchecked verdict.
A Bulletproofs proof can show that a decision's score was within policy bounds without revealing the score — useful when sharing evidence with a third party.
Every ledger entry is signed with both Ed25519 and ML-DSA-65 (FIPS 204), so records stay verifiable if classical signatures weaken.
VOPRF pseudonyms, private sanctions screening and SEAL encrypted scoring let institutions cooperate without handing each other raw customer data.
QUORUM is pre-pilot. Detection accuracy has not yet been measured on labelled data and no independent review has yet been performed; both are the next steps. What can be checked today is how it decides and how each decision is recorded.