System Governance
QUORUM separates who can change the rules from who uses them. New rules are observed before they act, sensitive changes need more than one person, and every change is written to the signed audit ledger.
Rule Lifecycle
Define
Each institution has its own rule set (plus optional global rules). Rules referring to unknown facts or operators are rejected when loaded rather than failing at decision time.
Shadow
New rules — including those proposed by the AI red-team tools — run in shadow mode alongside live traffic. QUORUM records when they would have fired; they do not affect verdicts.
Approve
Promotion is an operator action with WebAuthn step-up. Where the institution requires it, promotion also needs a FROST threshold approval (RFC 9591) that names the exact rule and can be used only once.
Promote & Record
The rule set is updated and kept in sync across instances. The change is written to the hash-chained, signed audit ledger.
Who Can Do What
Operators
RBACRole-based access (platform_admin, security_admin) to the operator console and API.
Privileged actions
WEBAUTHNStep-up authentication with a hardware or platform authenticator before sensitive changes.
Multi-party approvals
FROSTThreshold approvals with distributed key generation for rule promotion and user reinstatement. No single person can produce the approval.
Regulatory reports
DUAL CONTROLA SAR/STR prepared by one person must be approved by a different person. Nothing is filed automatically.
Role of the AI Models
Three locally run, fine-tuned models (Sentinel, Inquisitor and Adversarial) analyse cases asynchronously. When they disagree, or risk is high, an arbitrator weighs their reasoning; if it is unavailable, a recorded 2-of-3 majority applies. Calls use temperature 0, which reduces variation but does not guarantee identical output on different hardware.
The models never issue the real-time verdict and cannot change rules on their own. If they are unavailable, the rule engine is unaffected; deep-analysis requests return "challenge" (or "block" if the institution chooses fail-closed).
Explanations
Every real-time block gets a signed counterfactual certificate stored against its decision ID: each fired rule's contribution, the score and verdict had that rule not fired, and the smallest set of rules whose absence would change the verdict. Certificates are issued off the response path, so a failure never changes the decision.
These certificates are supporting evidence for explanation obligations such as EU AI Act Article 86 or adverse-action notices. Whether they meet a particular obligation is for the institution and its advisers to decide.