System Overview

Platform

QUORUM is a governed risk decisioning platform. It operates as a high-integrity decisioning layer that sits between event ingestion and institutional ledger enforcement — combining cryptographic attestability, governed adaptation, and forensic continuity in a single architecture.

Pipeline Position: Decision Engine

01 // Decisioning Engine

The core engine evaluates every incoming event through a three-tier analysis plane: deterministic rule evaluation, heuristic behavioral scoring, and consensus arbitration. Every decision output is cryptographically signed and committed to the audit ledger before enforcement.

  • Multi-dimensional signal ingestion with schema enforcement
  • Governed rule arbitration with policy precedence
  • Cryptographically attested, explainable decision outputs
  • FROST(Ed25519) threshold signing — no single key compromise can forge an audit verdict
[ENGINE_INIT]
Tier 1: SENTINEL (Deterministic Rule Eval) ACTIVE
Tier 2: INQUISITOR (Behavioral Scoring) ACTIVE
Tier 3: Adaptive Rule Engine (Consensus) ACTIVE
FROST Threshold Signing ACTIVE
KZG Audit Proofs (O(1) inclusion) ACTIVE
// ALL SYSTEMS OPERATIONAL

02 // Risk Scoring Pipeline

Ingestion

Raw event data is normalized and hydrated with historical context. Postgres WAL provides crash-consistent event buffering.

Feature Extraction

Synthesizing transaction velocity, entity age, and behavioral drift. Institutions apply mandatory local differential privacy before submission.

Governed Evaluation

Multi-tier analysis with pairwise secure gradient aggregation — server sees only the weighted gradient sum, never individual institution gradients.

Attested Verdict

The Adaptive Rule Engine issues a final score and enforcement directive, signed via FROST threshold protocol and committed to the KZG audit ledger.

03 // Behavioral Correlation Layer

Entropy-Based Identification

Instead of relying on static identifiers, QUORUM analyzes the 'entropy' of user behavior—keyboard rhythms, navigation paths, and transaction timing—to create a characteristically unique fingerprint.

DETECTED_DRIFT: 0.142 (THRESHOLD: 0.15)
FINGERPRINT_STABILITY: 98.2%

Cross-Event Mapping

The correlation layer maps relationships across seemingly unrelated events to identify coordinated fraud rings and automated attacks before they manifest as a loss.

04 // Rule Arbitration & Governed Adaptation

Logic Precedence

QUORUM's Adaptive Rule Engine resolves conflicts between human-defined rules and model-generated scores. Institutional mandates take precedence. Rules undergo shadow evaluation against historical event streams before promotion — no rule enters production without quantified impact.

Every rule change requires cryptographic authorization from designated compliance officers. Deployment is atomic and fully traceable in the append-only audit ledger.

STRICT_MODE RULE > MODEL
ADAPTIVE_MODE MODEL > RULE (Within Params)
BALANCED_MODE CONSENSUS REQUIRED
SHADOW_EVAL REQUIRED BEFORE PROMOTION
SIGNING_QUORUM FROST (t,n) THRESHOLD

05 // Human-in-the-loop Constraints

While QUORUM is designed for autonomy, institutions can enforce strict human-intervention gates for high-value or high-risk decisions.

Review Thresholds

Automatically escalate decisions falling within a specific 'Ambiguity Window' to human analysts.

Manual Overrides

Grant authorized officers the ability to override system verdicts with mandatory reason-code logging.

06 // Decision Output Types

BLOCK

Immediate enforcement. Transaction terminated.

ALLOW

Risk within acceptable parameters. Continue.

REVIEW

Pending human validation. Held for 30s.

ESCALATE

High-value exception. Level 3 clearance required.

04 // Live Engine Preview

Watch the analysis pipeline process an event in real time — from ingestion through AI consensus to verdict enforcement.

QUORUM — Live Analysis Engine
IDLE
Launch Full Engine
T1 — Sentinel
T2 — Inquisitor
T3 — Adversarial
Composite Score

06 // Cryptographic Infrastructure

Every decision is cryptographically attested at the moment it is made. The following primitives form the trust stack that makes QUORUM decisions provably correct and independently verifiable.

Threshold Signing
FROST(Ed25519)
(t,n) threshold Schnorr signing. No single key compromise can forge an audit verdict. Institution holds co-signing authority.
Audit Proofs
KZG Polynomial
O(1) inclusion proofs — 48-byte G1 point per audit entry. Any party verifies in two pairings. Replaces O(log n) Merkle traversal.
Privacy Accounting
Rényi DP (RDP)
Tight moments accountant with HSM-signed privacy certificates. Institutions prove ε/δ guarantees to regulators with a verifiable artifact.
Sanctions Screening
PIR / VOPRF-PSI
Ristretto255 VOPRF. The screening service learns only that some entity was checked — the queried name never appears in cleartext.
Federated Learning
Pairwise Secure Gradient Aggregation
X25519 ECDH masks cancel at the aggregation server. The server sees only the weighted gradient sum — never individual institution gradients. Mandatory local DP applied before submission; server verifies but never sees raw gradients.
Explainability
Causal Counterfactual Certificates
HSM-signed per-decision explanations satisfying EU AI Act Article 86 and FINTRAC adverse-action notice obligations. Generated at decision time — not reconstructed after the fact.

05 // Architecture Documentation

Detailed specification for each subsystem in the QUORUM pipeline. Each document covers data contracts, processing logic, failure modes, and integration surface.

Full architecture specification → architecture.html