Risk Scoring API
Real-time fraud risk assessment that fuses 151 detection techniques across 12 independent analyzers through a Beta distribution conjugate prior with calibrated posterior inference. One endpoint, one mathematically grounded score, one explainable decision — with 47ms p95 for the fast path, with full Bayesian uncertainty quantification.
No opaque neural networks. Every detection layer produces an explainable signal with a traceable scientific basis. The Fusion Core accumulates evidence through conjugate Bayesian updating — not weighted averages, not gradient-descent black boxes. Every decision is mathematically auditable.
Live Risk Analysis
Dual-Path Decision Architecture
Every inbound request enters a triage subsystem that routes unambiguous decisions through the O(1) fast path and reserves the full Bayesian detection stack for cases in the statistical grey zone — optimizing both latency and detection depth.
Fast Path
47ms p95Resolves 85%+ of all requests with O(1) computational complexity. Employs Bloom filter membership queries against known bad-actor fingerprint sets, static and dynamic bot signature matching, basic velocity checks via edge KV store, and user-agent heuristic classification.
Slow Path
<200msReserved for ~5% of traffic in the statistical grey zone where fast-path heuristics are inconclusive. Activates the full 151-technique detection stack across 12 analyzers with Bayesian posterior inference, nine information-theoretic entropy measures, propensity score causal inference, and game-theoretic decision boundary certification.
Detection Layer Evidence Weights
Each detection layer produces an independent risk signal with its own calibrated confidence interval. The Fusion Core accumulates evidence through the Beta distribution as a conjugate prior for Bernoulli-distributed fraud indicators — not a naïve weighted average, but proper Bayesian posterior updating with uncertainty propagation.
Device Fingerprint
Canvas spectral entropy analysis, GPU shader compilation timing, anti-detect browser topology detection (13+ commercial tools), WebAuthn AAGUID hardware attestation verification, and SimHash cross-session identity resolution
Behavioral Biometrics
Physiological micro-tremor detection (8–12 Hz band), Hurst exponent estimation via R/S analysis, jerk kinematic profiling, keystroke di-graph temporal patterns, and curvature consistency metrics
IP Intelligence
ASN reputation scoring, Haversine geo-velocity computation (impossible travel >900 km/h), VPN/Tor/proxy infrastructure detection, residential vs. datacenter IP classification, and hosting provider taxonomy
Threat Consortium
Privacy-preserving consortium lookup via Bloom filter syndication (1M entries, ~1% FP) and differentially private signal sharing (Laplace mechanism, ε=0.5) with HMAC-SHA256 signed contributions
Velocity & Rate
Exponential moving average (EMA) adaptive rate smoothing, burst pattern detection via Z-score anomaly analysis (Welford online algorithm), and sliding window velocity profiling across configurable temporal granularities
Session Anomaly
Markov chain state transition modeling with impossible-transition detection, CUSUM sequential change-point analysis (Page, 1954), fingerprint drift tracking, and temporal coherence validation
Bayesian Fusion Core
The mathematical engine behind every risk decision. Evidence accumulation via the Beta distribution conjugate prior for Bernoulli-distributed fraud signals — yielding closed-form posterior updates, calibrated uncertainty, and computationally efficient online inference without MCMC sampling overhead.
Beta(α₀, β₀) = Beta(1, 1)Uniform prior — each request starts with no assumption about legitimacy.
α += Σ wᵢ · (1 - sᵢ) · cᵢ | β += Σ wᵢ · sᵢ · cᵢEach detection layer contributes a score sᵢ (0-1) with confidence cᵢ (0-1), weighted by importance wᵢ. High fraud probability increases β; low increases α.
P(legitimate) = α / (α + β)The mean of the posterior Beta distribution gives the probability of legitimacy.
P(fraud) = 1 - α / (α + β)The complement: probability that this request is fraudulent.
Var = αβ / ((α+β)²(α+β+1))High variance → low confidence → may trigger challenge instead of binary decision.
API Response Schema
Every response includes the actionable decision, per-layer explainable risk signals, detection flags, calibrated confidence breakdown with data coverage and model calibration metrics, and a cryptographically signed evidence chain for forensic auditability.
POST /api/v1/analyze → 200 OK
{
"requestId": "req_a7f2c9b1",
"visitorId": "vid_x8k3m2n9",
"deviceHash": "sha256:3f7a...",
"decision": "challenge",
"riskScore": 0.42,
"confidence": 0.87,
"confidenceBreakdown": {
"dataCoverage": 0.92,
"modelCalibration": 0.85,
"systemMode": "normal"
},
"entropyScore": 0.71,
"entropyLevel": "MEDIUM",
"flags": {
"isBot": false,
"isIncognito": true,
"isEmulator": false,
"isVM": false,
"isAutomation": false
},
"reasons": ["new_device", "incognito_detected"],
"signals": {
"deviceFingerprint": { "score": 0.12, "confidence": 0.95 },
"behavioral": { "score": 0.08, "humanLikelihood": 0.72 },
"ipIntel": { "score": 0.10, "isProxy": false },
"consortium": { "score": 0.07, "priorFraud": false },
"velocity": { "score": 0.03, "requestsLastHour": 12 },
"session": { "score": 0.02, "anomalyDetected": false }
},
"processingTimeMs": 4,
"evidenceId": "ev_9c2d..."
}Asymmetric Positive Trust Model
Not all evidence is symmetric. Only designated trust-lowering layers can actively reduce the posterior fraud probability — all other layers can only increase risk or remain neutral. This game-theoretic constraint prevents adversaries from accumulating counterfeit trust signals to manufacture a low-risk score through strategic manipulation.
Trust-Lowering Layers
These layers can reduce the fraud score when evidence of legitimacy is present.
Risk-Only Layers
These layers can only raise or maintain the current fraud score — never lower it.
Unified Risk Intelligence — One Endpoint, Full Explainability
A single API endpoint fusing 151 detection techniques across 12 analyzers into one actionable, mathematically grounded score — with full Bayesian explainability, calibrated confidence intervals, and 47ms p95 fast-path latency at the edge.