Risk Scoring Methodology

ForensicBlock's risk scoring model is designed to be transparent, reproducible, and legally defensible. Every score can be traced back to specific on-chain data and verifiable risk factors.

Model version: v1.7.4 | Last updated: July 2026

Model Overview

The ForensicBlock risk score is a tiered, evidence-weighted composite of independent risk signals. Every finding is classified by evidence tier — HARD (attributed, anchored evidence), STRONG (corroborated behavioral signals), SOFT (uncorroborated indicators) — before any aggregation. HARD evidence floors the verdict at CRITICAL; SOFT evidence alone can never push a score into HIGH. The model uses no black-box components — every factor, threshold, and floor is documented here and versioned.

450+
Curated active entities
2,300+
Address attributions with source + confidence
100%
Deterministic scoring

Risk Factors & Tiers

Scoring is tier-ordered, not flat-weighted: entity attribution overrides, exposure analysis is primary, graph and behavioral signals are supplementary context. The final score is reconciled so a lower-tier signal can never dilute higher-tier evidence.

Entity Attribution (Tier 1)

Override

Verified ground-truth entity match. When an address matches a known entity in our curated database (450+ active entities including OFAC SDN, state actors, known hackers, with 2,300+ address attributions), the entity's risk score is used directly. Entity attribution always overrides all other scoring methods. Every attribution carries source and explicit confidence.

Data source: ForensicBlock Intelligence Database, OFAC SDN List

Counterparty Exposure Analysis (Tier 2)

Primary

Deterministic scoring based on counterparty transaction exposure. For each transaction, the counterparty address is categorized (sanctions, darknet, mixer, fraud, exchange, DeFi, unknown) and exposure contributes by category weight and confidence. As of methodology v1.7.4, sanctioned exposure floors the verdict at CRITICAL only when it is SUBJECT-LINKED — a directed fund-flow path between the subject and the sanctioned counterparty within the traced graph. Sanctioned entities present elsewhere in the graph render as contextual intelligence with hop distance, never as a verdict override.

Data source: Graph analysis, pre-computed exposure_analysis, entity database lookups

Graph Risk Score (Tier 3)

Supplementary

Risk derived from the full transaction graph traversal — node-level risk aggregation, high-risk path detection, and cluster analysis. Used when exposure analysis data is not yet available or as supplementary context.

Data source: ForensicBlock Graph Engine, Alchemy Asset Transfers API

Behavioral Analysis (Tier 4)

Supplementary

AI agent-based behavioral scoring — transaction velocity, pattern detection, anomaly analysis. Used as supplementary context when entity and exposure data are unavailable. The MAX of all tiers is used as the final score — behavioral analysis never dilutes a higher-tier score.

Data source: ForensicBlock AI Agent Pipeline (6 agents)

Risk Categories (18 FATF-aligned)

Varies

Sanctions/OFAC (weight 1.0), Terrorist Financing (1.0), State Actor (1.0), Darknet Market (0.85), Ransomware (0.85), Stolen Funds (0.85), Mixer/Tumbler (0.80), Fraud/Scam (0.80), Cybercrime (0.80), High-Risk Exchange (0.55), P2P Exchange (0.45), Gambling (0.35), Unregulated Service (0.40), Cross-chain Bridge (0.30), Regulated Exchange (0.05), DeFi Protocol (0.10), NFT Marketplace (0.05), Unknown (0.15).

Data source: lib/risk-categories.ts — FATF typology framework

Score Thresholds

—

LOW: 0–39. MEDIUM: 40–69. HIGH: 70–89. CRITICAL: 90–100. Same address + same blockchain state = identical score every time. Methodology version: v1.7.4.

Data source: Deterministic computation — reproducible and court-defensible

Token Footprint

Supplementary

Number of distinct token types held or transacted. Unusually high token diversity may indicate DeFi farming, airdrop harvesting, or wash trading.

Data source: Alchemy getTokenBalances

Address Age

Supplementary

Time since the address's first on-chain activity. Addresses less than 30 days old receive elevated risk scores (up to 70 for brand-new addresses) as they are more likely to be disposable addresses used in laundering schemes.

Data source: First transaction timestamp

Score Calculation

score = baseline; HARD evidence → floor at 90 (+2 per additional HARD factor, max +10); STRONG factors → saturating sum (cap 70); SOFT factors → log-saturating sum (cap 35); level = canonical band lookup

The caps are chosen so evidence tiers cannot impersonate each other: STRONG signals alone can reach HIGH but never CRITICAL (CRITICAL requires HARD, attributed evidence), and SOFT signals alone top out at the bottom of MEDIUM. Verdict floors (sanctions, attributed mixer, illicit-entity) apply only to subject-linked exposure — see the v1.7.4 changelog entry below. Every scored result carries its full derivation: base score, each factor, each floor that fired, and which subject-linked address triggered it.

Confidence Score

Every risk assessment includes a confidence score (0.20–0.95) reflecting the completeness and quality of available data:

  • Base confidence starts at 0.30 and increases with each available signal (up to 8 signals = +0.60)
  • Addresses with fewer than 3 transactions receive a -0.10 confidence penalty
  • During multi-agent investigations, confidence is further adjusted by the verification rate of findings

Risk Level Thresholds

One scale governs every surface — this page, the product UI, and the sealed report verdict all map scores through the same published bands (engine constant TIER_THRESHOLDS, unchanged since v1.5.0 so historical scores stay comparable).

0–39
Low

No significant risk indicators. Standard due diligence sufficient. Scores 0–9 carry no elevated risk indicators at all.

40–69
Medium

Some risk indicators present. Enhanced due diligence recommended.

70–89
High

Multiple corroborated risk indicators. Investigation and compliance review required.

90–100
Critical

Severe risk backed by hard evidence — subject-linked sanctions exposure, attributed mixer usage, or confirmed illicit activity.

Multi-Agent Validation

During full investigations, ForensicBlock employs six specialized AI agents. Each agent produces independent findings that are cross-validated by the orchestrator:

  • TRACER — Multi-hop fund-flow tracing with bridge and mixer detection
  • SENTINEL — OFAC/SDN screening and per-factor risk scoring
  • EXAMINER — Behavioral pattern analysis — peel chains, fan-out, layering, anomalies
  • HUNTER — Recovery probability assessment and exchange identification
  • CUSTODIAN — Real-time address monitoring and custody-relevant event capture
  • SEALER — Evidence compilation and sealed, court-ready report generation

Findings must be verified against on-chain evidence before being included in the final report. The overall confidence score is reduced proportionally to unverified findings: adjusted = confidence * (0.5 + 0.5 * verificationRate)

Data Sources

Alchemy

Primary blockchain data provider. Real-time transaction data, asset transfers, token balances, and webhook-based monitoring across EVM chains.

Etherscan V2

Secondary provider for historical transaction data. Multi-chain support via chain ID parameter.

OFAC SDN List

Official U.S. Treasury Specially Designated Nationals list. Updated regularly and cached locally with 5-minute refresh cycles.

ForensicBlock Intelligence

Curated database of 450+ active entities and 2,300+ address attributions covering exchanges, DeFi protocols, bridges, mixers, scams, and sanctioned entities — every attribution with source and explicit confidence.

Reproducibility & Defensibility

ForensicBlock is designed for legal proceedings:

  • Deterministic scoring — The same inputs always produce the same score. No randomness in the risk model.
  • Evidence citations — Every risk factor cites specific transaction hashes, block numbers, or data source references.
  • SHA-256 integrity — Generated PDF reports include a SHA-256 hash of the complete document stored on record for tamper detection.
  • Audit trail — Every agent action, finding, and decision is logged with timestamps for full investigative chain of custody.
  • Model versioning — Risk scores include the model version used, enabling reproduction of historical assessments.

Methodology versions

Methodology changes ship as new, dated versions. A sealed report permanently cites the version (and its SHA-256 content hash) it was generated under — prior sealed reports are never retroactively re-scored or re-versioned. That discipline is what makes the version stamp on a report mean something.

v1.7.42026-07-05active

Affirmative-evidence text classification. Text-pattern evidence-tier classification requires affirmative, non-negated language: a finding whose prose negates the matched term (for example, "no sanctioned addresses identified") is treated as a disclosure of a clean result, not as a hit. High-tier text classification additionally requires the finding to carry high or critical severity. Typed evidence attribution (entity category, factor tier) remains the authoritative classification path and is unaffected.

v1.7.32026-07-05

Evidence-grade linkage. The subject-linkage fund-flow analysis traverses native-value edges only; token-denominated transfers contribute undirected graph-proximity context but never a verdict floor. When a traced graph is evaluated, attributions that cannot be located within it are treated as contextual intelligence rather than verdict drivers. A conservative unscoped mode applies, and is disclosed, when no traced graph is available.

v1.7.22026-07-02

Verdict floors (sanctions / mixer / illicit-entity) now apply only to SUBJECT-LINKED exposure — attributions with a directed fund-flow path to or from the investigation subject. Peripheral attributions elsewhere in the traced graph render as contextual intelligence with hop distance, never as a verdict override. Published risk bands reconciled to the engine's canonical thresholds (Low 0–39, Medium 40–69, High 70–89, Critical 90–100).

v1.7.12026-06-12

Finding hygiene ahead of scoring: pattern-class deduplication (corroboration raises confidence, not count), unanchored findings excluded from scoring, unattributed structural/mixer findings capped at STRONG — "potential X" never counts as "X confirmed".

v1.7.02026-03-15

Tiered evidence-weighted scoring: every finding classified HARD / STRONG / SOFT before aggregation. HARD evidence floors the verdict at SEVERE; SOFT evidence alone can never reach HIGH. Methodology version + SHA-256 content hash stamped on every scored result.

FRE 902(13) & 902(14) — what we do, what we don't

Federal Rules of Evidence 902(13) and 902(14), in effect since 12/01/2017, allow electronic records to self-authenticate when accompanied by a qualified person's written certification. Rule 902(13) covers records generated by an electronic process; Rule 902(14) covers data copied from an electronic device. Both rules eliminate the need for a live custodian to authenticate at trial.

What our sealed reports carry:

  • The methodology version (v1.7.4 today) plus the SHA-256 content hash of that methodology — independently re-derivable.
  • The anchored evidence SHA-256s of every finding the report relies on.
  • The Bates range, atomically allocated (one range, no holes).
  • A chain-of-custody log whose entries are hash-chained.
  • The qualified-person certification block — examiner of record, training, methodology version, anchored hashes.
  • A public verifier URL anyone can open in a browser without an account.

What 902(13)/(14) does NOT do — the over-claim we never make:

Self-authentication is not the same as admissibility. The fact that a record self-authenticates means a court will accept that it is what it claims to be — a record generated by our electronic process. It does not resolve hearsay (FRE 801–807), relevance (FRE 401–403), or the Daubert qualification of the underlying methodology (FRE 702). Those remain separate hurdles your counsel argues independently. ForensicBlock packages for self-authentication; we never represent that a sealed report is "automatically admissible."

Daubert posture — built for cross-examination

In U.S. v. Sterlingov (2024, D.D.C.), the court admitted deterministic Bitcoin clustering under FRE 702 — but put the proprietary methodology behind it under scrutiny that the prevailing label-only tools did not enjoy. ForensicBlock was designed for exactly that scrutiny.

The Daubert factors, mapped to ForensicBlock:

  • Tested: deterministic — same input, same output, every run.
  • Peer review: methodology is public on this page, versioned, content-hashed.
  • Known error rate: every entity attribution carries an explicit confidence.
  • Standards & controls: the SQL classifier (fb_classify_onchain_value), the canonical finalizer (fb_finalize_investigation), the audit log chain.
  • General acceptance: built on the same evidentiary primitives (deterministic clustering, glass-box attribution) the broader forensic field accepts.

What we don't carry to court:

  • Black-box risk scores. Every score decomposes into source + confidence per row.
  • Probabilistic mixer re-attribution dressed up as certainty. Mixer ingress is stated as a pause.
  • Privacy-coin tracing where the technology doesn't support it. We say so on the report's face.
  • Recovery guarantees. Recovery is a legal action. We do the evidence.

The court-ready report spec — 32 sections, in order

Every sealed court-ready report carries these 32 sections, in this order. This list is rendered from the same template definition the report generator executes — the published spec and the shipped report cannot drift apart. Standardization is part of the moat: opposing counsel can verify against a fixed spec; expert testimony lands on the same scaffolding every time. A section with no underlying data still appears, explicitly marked as unpopulated with the reason — honest empty, never a silent omission.

  1. 01Cover & Seal Identifiers
  2. 02Table of Contents
  3. 03Executive Summary
  4. 04Subject Profile
  5. 05Scope & Methodology
  6. 06Findings
  7. 07Fund Flow
  8. 08Transaction Statistics
  9. 09Fund Landing & Recovery
  10. 10Transaction Graph (Topology)
  11. 11Graph Topology Metrics
  12. 12Sanctions & Counterparties
  13. 13Behavioral Indicators
  14. 14Risk Assessment
  15. 15Evidence Inventory
  16. 16AI Analysis (Machine-Generated Narrative)
  17. 17Examiner Notes
  18. 18Chain of Custody
  19. 19Admissibility Checklist (FRE 901 / 902(13) / 902(14))
  20. 20Legal Certifications
  21. 21Explicit Limitations
  22. 22Conclusions
  23. 23Expert Witness Scaffolding (FRE 702)
  24. 24Screening History
  25. 25Entity Attributions
  26. 26Exhibit Index
  27. 27Address & Transaction Index
  28. 28Independent Verification Walkthrough
  29. 29Appendix A — Evidence Anchor Hashes
  30. 30Appendix B — Methodology & Content Hash
  31. 31Appendix C — Methodology Change History
  32. 32Appendix D — Glossary

The Audit Pack — verify our work, end-to-end

Every investigation can produce a single downloadable JSON envelope — the Audit Pack — that carries everything a defense expert, opposing counsel, or regulator needs to reproduce, challenge, or independently verify the report. The pack is sealed under a self-hash so any tampering is immediately detectable.

What the pack contains:

  • The methodology version + content hash
  • The input snapshot (seed address, chain, depth)
  • Every agent execution — input, output, status, timing
  • Every finding — with anchored evidence SHA-256
  • The hash-chained audit log (tamper-evident)
  • The graph payload (nodes + edges, methodology-stamped)
  • The behavioral-risk engine envelope
  • The value-quality classifier envelope
  • Re-derivation instructions in plain English
  • A self-hash binding the pack to itself

Why this is the moat:

What the proprietary incumbents cannot ship: attribution you can republish in an exhibit (their label databases are license-locked), a public verifier opposing counsel can open without an account, and an exportable, self-hashed Audit Pack that lets a defense expert re-derive the work. We win by going the other way: every methodology decision is public on this page, every input recorded, every output exportable, every audit-log entry hash-chained. Defense experts who try to invalidate our reports end up re-deriving the same conclusions — and that is precisely the surface our sealed certifications stand on.

The pack is available per investigation from the investigator's detail page ("Quick Actions → Audit pack") and from the report detail page once a report has been sealed. Authenticated users get the pack scoped to their own matters; for public audit of a specific sealed report, opposing counsel can request the pack from the issuing firm or use the public verifier.

Ready to investigate?

Try a free risk check on any blockchain address, or sign up to run full forensic investigations with all six AI agents.