System Overview

Platform

QUORUM is a fraud and risk decisioning platform for regulated financial institutions, deployed inside the institution's own environment. It runs alongside existing fraud and case-management systems: it returns a verdict, a risk score, the rules and signals behind it and a decision identifier, and the institution's systems decide what to do.

Pipeline Position: Decision Engine

01 // Decisioning Engine

Real-time verdicts come from a per-institution rule engine combined with behavioural scoring. Each verdict is recorded before it is returned and sealed within seconds into a hash-chained ledger signed with Ed25519 and ML-DSA-65. If the rules cannot be evaluated or the decision cannot be recorded, no verdict is issued.

  • REST, ISO 8583, message-queue and flat-file decision channels
  • Facts the request does not carry are treated as absent — no invented defaults
  • Fired rules, signals and a decision ID returned with every verdict
  • Fail closed: HTTP 503 / ISO 8583 code 91 rather than a silent "allow"
[DECISION_PATH]
Rule engine (per institution) REAL TIME
Behavioural scoring (statistical) REAL TIME
Decision record (hashed before release) REAL TIME
Ledger seal (Ed25519 + ML-DSA-65) SECONDS
AI tiers: Sentinel · Inquisitor · Adversarial ASYNC REVIEW
// AI output is evidence for review, never on the real-time path

02 // Risk Scoring Pipeline

Intake

Each channel maps its input to rule facts and leaves out anything the input does not carry. Requests are scoped to the calling institution.

Facts & Velocity

Velocity per card or account over 1 hour and 24 hours is counted from the decision log, keyed by an institution-scoped HMAC — never the raw PAN.

Evaluation

Behavioural score (60%) combined with the rule engine (40%) when transaction facts are supplied. If either alone reaches 70, the verdict is at least ELEVATED_RISK.

Recorded Verdict

A canonical record holding a digest of the inputs is hashed and stored before the verdict is returned, then sealed under a Merkle root into the signed ledger.

03 // Behavioural & Device Signals

Interaction Signals

A browser SDK reduces typing rhythm, pointer movement, touch and motion data to derived values. Per-user baselines (Welford's running mean and variance) are kept for each institution, and deviations raise the behavioural score.

SIGNALS: typing speed & variance · pointer entropy · touch force · frame jitter
AUTOMATION: headless / webdriver · UA vs TLS (JA3) mismatch

Relationship Graph

Where an institution enables it, an optional Neo4j graph links accounts, devices and IP addresses so that risk on one entity can raise the risk of entities connected to it. No measured uniqueness rate is claimed for device fingerprints.

04 // Rule Governance

Shadow First

Each institution has its own rule set, validated when loaded. New rules — including any proposed by the AI red-team tools — start in shadow mode, where they are evaluated alongside live traffic without affecting verdicts.

Promotion is an operator action. 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. Every change is written to the signed audit ledger.

REAL-TIME VERDICT RULES + BEHAVIOURAL SCORE
AI MODELS ASYNC REVIEW ONLY
NEW RULES SHADOW MODE
PROMOTION OPERATOR (+ FROST WHERE REQUIRED)
PRIVILEGED ACTIONS WEBAUTHN STEP-UP

05 // Human Authority

QUORUM is additive: the institution's own systems keep final authority over approval, decline, step-up and review. Pilots run in shadow mode before QUORUM is given any production authority.

AI Findings for Analysts

Three locally run AI tiers review cases asynchronously. When they disagree, an arbitrator weighs their reasoning. The result goes to analysts as evidence; it does not change a verdict by itself.

Regulatory Reports

High-risk blocks open a draft SAR/STR case. A person completes it and a different person approves it. Nothing is filed automatically.

06 // Verdicts

VERIFIED

Score below 15.

AUTHORIZED

Score 15–39.

ELEVATED_RISK

Score 40–69. Recorded as flagged for the institution's review process.

ISOLATED

Score 70 or more. Rejected on the transaction channel; a signed counterfactual certificate is issued.

08 // Illustrative Analysis Preview

An illustrative animation with sample values, not output from the engine. In QUORUM the real-time verdict comes from the rule engine and behavioural score; the three AI tiers shown here review cases afterwards and their findings go to analysts.

QUORUM — Illustrative Preview
IDLE
Open Simulator
T1 — Sentinel
—
T2 — Inquisitor
—
T3 — Adversarial
—
Composite Score
—

07 // Cryptographic Infrastructure

The mechanisms below make each recorded decision tamper-evident and provable to a third party. The ledger is tamper-evident rather than immutable: it lives in PostgreSQL, but any change to an anchored entry is detectable without trusting QUORUM.

Ledger Signatures
Ed25519 + ML-DSA-65
Each ledger entry carries the previous entry's hash and both a classical and a post-quantum (FIPS 204) signature; both must verify. A scan every 15 minutes alerts on any broken link.
Inclusion Proofs
Merkle Trees
A short proof that a decision is under a published root, checkable with SHA-256 alone. Roots can be timestamped by an RFC 3161 authority.
Zero-Knowledge
Bulletproofs
Proof that a decision's score was within stated bounds without revealing the score. No trusted setup; verifiable offline.
Approvals
FROST (RFC 9591)
Threshold Schnorr approvals with distributed key generation for sensitive actions such as rule promotion. Each approval names one action and target and is used once.
Federation & Sanctions
VOPRF (RFC 9497)
Institutions derive matching pseudonyms for the same card or device without revealing it, and screen names against a sanctions token set without QUORUM seeing the names or results. Exact match after normalisation only.
Encrypted Scoring
Microsoft SEAL (BFV)
A behavioural risk score computed on the institution's ciphertext; only the institution can decrypt. Federated model updates carry differential-privacy noise added by the institution, with a per-institution privacy budget.
Explainability
Counterfactual Certificates
Every real-time block gets a signed certificate: each fired rule's contribution and the smallest set of rules whose absence would change the verdict. Supporting evidence for explanation obligations — not a compliance determination.

09 // Architecture Documentation

A short description of each subsystem in the QUORUM pipeline. The full technical reference (QRM-ARCH-001 Rev 2.0) is available to institutions under NDA.

Full architecture specification → architecture.html