Model Card

Public documentation of the Riskora scoring model — architecture, training data, calibration, and measured performance. Last model update: July 2026.

Security & Compliance

0.983

AUROC

2.8%

FP rate @≥25

32K+

Labeled wallets

Model overview

Gradient boosting classifier

Histogram-based gradient boosting (HGB) over a compact set of on-chain behavioral features. CPU inference in milliseconds. Trained July 2026.

Isotonic probability calibration

Raw model output is calibrated with isotonic regression on a held-out split, then prior-shifted to a conservative 5% assumed illicit prevalence — scores reflect production-like conditions, not the training distribution.

Deterministic by design

Same on-chain state, same score. A fully rule-based mode (mode=rule) is available when maximum reproducibility is required, and every scan is stored as a citable audit record.

Human oversight

Scores are decision-support indicators for compliance analysts. They are not proof of illicit activity and must not be the sole basis for blocking, freezing, or reporting decisions.

Scoring pipeline

Four layered stages. Every score can be decomposed into the evidence each stage contributed.

1List screening

OFAC SDN, curated threat-intelligence blacklists, and a whitelist of known services (exchanges, pools, payment processors). Direct matches short-circuit the pipeline.

2Behavioral model

The HGB classifier scores 14 features covering activity, value flow, and transaction structure — calibrated to a probability of illicit use.

3Graph taint overlay

Multi-hop tracing surfaces mixer, coinjoin, and blacklist exposure that behavioral features alone cannot capture. Trace depth adapts to the wallet's risk profile.

4Whitelist normalization

Exposure to known legitimate services deflates the score proportionally — an exchange-heavy counterparty graph is not, by itself, evidence of risk.

Model features

A compact set of behavioral features computed from public on-chain data — no off-chain identity data is used. Feature definitions are proprietary and are not publicly disclosed.

Activity

How the wallet transacts over time — cadence, recency, and throughput patterns that distinguish services from individuals and illicit operators from both.

Value flow

How funds move through the wallet — accumulation, turnover, and spending behavior relative to its apparent purpose.

Transaction structure

How transactions are constructed — input and output topology, fee behavior, and consolidation patterns typical of mixing and peel chains.

Training data

Approximately 32,000 labeled Bitcoin addresses, split 60/20/20 into fit, calibration, and test partitions.

15,011Decontaminated on-chain normals

Randomly sampled real Bitcoin wallets, screened against all blocklists before labeling — the 'ordinary user' class that prevents service-heavy false positives.

5,000Exchange wallets

Known exchange/service addresses, teaching the model legitimate high-volume behavior.

9,840Stratified illicit classes

Darknet markets (3,958), tumblers/mixers (3,432), blackmail & extortion (2,434), cybercrime services (1,821), plus ponzi schemes, money laundering, and government-blacklisted addresses — from the BABD-13 academic dataset.

Held-out performance

Measured on a test partition of 6,435 labeled addresses never seen during training or calibration. AUROC 0.9834 · average precision 0.9535.

Score threshold	Typical use	False-positive rate (licit)	Recall (illicit)	Precision
≥ 25	Standard review queue	2.8%	84.9%	93.0%
≥ 50	Priority review	1.4%	72.5%	95.8%
≥ 85	Critical alerts	0.3%	34.7%	97.9%

Higher thresholds trade recall for precision. The graph taint overlay adds multi-hop evidence on top of the base model score, so production detection coverage exceeds the model-only recall shown here.

Limitations & intended use

Scores are probabilistic risk indicators, not proof of illicit activity. A high score flags patterns consistent with illicit use; a low score is not a guarantee of legitimacy.

Performance figures are measured on a held-out labeled test set (n=6,435). Production traffic differs in composition; scores are prior-shifted to a conservative prevalence to compensate.

Wallets with very little on-chain history produce conservative scores — the graph taint overlay provides additional evidence where behavioral features are thin.

The model reflects patterns present in its training data. Novel laundering techniques are caught first by the list screening and graph overlay, then absorbed into periodic retraining.

Governance & compliance documentation

This model operates under a documented EU AI Act minimal-risk self-assessment. Scores are decision-support for compliance analysts — never the sole basis for a decision. DPA, TOMs, and our subprocessor list are public.

Security & Compliance Data Processing Agreement