Skip to main content

​
Graph Intelligence

Quint’s risk scoring is not a single model output. It’s a weighted composition of four signal layers, each producing a sub-score that contributes to a final 1-100 result. This makes every score fully explainable — you can decompose any score into its contributing signals. 
┌────────────────────────────────────────────────┐
│              Incoming Agent Event               │
└───────────────────────┬────────────────────────┘
                        │
                        ▼
          ┌─────────────────────┐
          │   Graph Builder     │
          │   (event → subgraph)│
          └──────────┬──────────┘
                     │
     ┌───────────────┼───────────────┐
     ▼               ▼               ▼
┌──────────┐  ┌──────────┐  ┌─────────────┐
│ Layer 1  │  │ Layer 2  │  │ Layer 3     │
│ Intrinsic│  │ Struct-  │  │ Policy      │
│ Action   │  │ ural     │  │ Violation   │
│ Risk     │  │ GNN      │  │ Score       │
└────┬─────┘  └────┬─────┘  └──────┬──────┘
     │              │               │
     ▼              ▼               ▼
┌─────────────────────────────────────────┐
│       Layer 4: Temporal Anomaly         │
│       Modifier (×0.5 to ×2.0)          │
└────────────────────┬────────────────────┘
                     │
                     ▼
          ┌─────────────────────┐
          │   Score Compositor  │
          │   final_score 1-100 │
          └─────────────────────┘

​
Architecture Components

Scoring Algorithm

4-layer composite scoring with configurable weights, confidence computation, and LLM fallback triggers

Forward-Chaining Engine

90 inference rules across 7 compliance frameworks with agenda-based fixed-point evaluation

Graph Neural Network

SubgraphGNN with 13 node types, 20 edge types, multi-task learning for risk classification

Graph-Structured RAG

Compliance articles and mitigations retrieved from Memgraph and injected into LLM prompts

Memgraph Co-Processor

Graph-native reasoning with centrality scoring, community detection, and impact propagation

Compliance Ontology

1,948 nodes, 1,075 edges across GDPR, HIPAA, SOC2, PCI-DSS, OWASP, ISO27001

Action Taxonomy

Canonical domain:scope:verb event classification with 7 domains and 150+ actions

​
The quint-graph Package

The graph intelligence layer is extracted as a standalone Python package: 
from quint_graph import GraphReasoner

reasoner = GraphReasoner()

result = reasoner.evaluate(
    event=event_data,
    policies=customer_policies,
    tenant_id="customer-uuid"
)

print(result.score)        # 1-100
print(result.risk_level)   # none | low | medium | high | critical
print(result.confidence)   # 0.0 - 1.0
print(result.violations)   # List of policy violations
print(result.fired_rules)  # Which inference rules fired

​
Key Stats

MetricValue
Inference rules90
Compliance frameworks7
Ontology nodes1,948
Ontology edges1,075
Evaluation time (forward-chaining)< 1ms
GNN node types13
GNN edge types20
Total hot-path latency< 50ms
LLM fallback rate~15-25% of events
Test suite194 tests

​
Feature Flags

All advanced capabilities are gated behind feature flags (default False):
FlagDefaultPurpose
graph_reasoner_enabledTruePrimary forward-chaining engine
memgraph_enabledFalseMemgraph enrichment co-processor
memgraph_enrich_enabledFalseCentrality-weighted enrichment
memgraph_event_store_enabledFalseWrite events to graph DB
gnn_enabledFalseSubgraphGNN structural scoring