Sanctions screening for compliance and risk

Watchman is an open-source screening engine you can inspect, tune, and defend in an exam. It downloads OFAC, EU, UK, UN, and related lists, scores each customer or counterparty against those lists, and returns a ranked hit with a field-level score — not a black-box yes/no.

This page is for BSA/AML officers, sanctions officers, model-risk reviewers, and internal audit. The technical scorer is described in Similarity methodology. Numbers below come from scoring the full OpenSanctions Pairs dump (755,540 analyst-labeled pairs, 293 sources, 45 jurisdictions) with production Similarity in September 2026.

The job screening has to do

Targeted financial sanctions require countries to freeze without delay the funds of designated persons and to stop those funds being made available to them (FATF Recommendations 6 and 7; FATF best practices on Recommendation 6). In the United States, OFAC administers those lists and evaluates apparent violations under the Economic Sanctions Enforcement Guidelines (31 CFR Part 501, Appendix A). OFAC’s Framework for OFAC Compliance Commitments (2019) expects a written program with management commitment, risk assessment, internal controls, testing, and training.

The Wolfsberg Group is explicit about what screening is: “the comparison of one string of text against another to detect similarities which would suggest a possible match,” used as one control inside a wider sanctions program, with a risk-based approach and known limits (Wolfsberg Guidance on Sanctions Screening, 2019). Screening is not CDD, not transaction monitoring, and not a substitute for investigating a true hit.

Two failure modes dominate operations:

1. A missed designation (false negative) can become a prohibited transaction. That is the “without delay” problem.
2. A flood of false hits (false positives) burns analyst hours, delays good customers, and hides real alerts in noise. Name-only tools are famous for this. OFAC’s own public search returned 21 matches at 100% for the query Khamis Al; Watchman’s name scorer on those same SDN names ranged from 0.26 to 0.87 (portal comparison).

A usable engine has to move both numbers, and it has to show why a hit scored the way it did. That last point is model risk: OCC/Fed SR 11-7 / OCC Bulletin 2011-12 expects conceptual soundness, ongoing monitoring, and effective challenge for models used in risk management.

What Watchman gives a program

Lists you can name. OFAC SDN, the US Consolidated Screening List, EU, UK, UN, FinCEN 311, plus CSV ingest of your own files. Refresh is periodic or on demand. You can point examiners at which file was loaded and when.

A score built from the fields CDD already collects. FinCEN’s CDD rule expects institutions to know who they are dealing with. Watchman takes that same record — name, type, government IDs, date of birth, address, contact, vessel IMO — and scores it as a structured entity, not a single string. Name-only queries are down-ranked on purpose.

How identifiers affect the score. Some identifiers are unique to one person or vessel. If the query and a list record share a passport, national ID, IMO number, MMSI, aircraft serial, or crypto address, and the country (where it applies) matches, Watchman returns 1.0 — treat this as the same party.

Other identifiers are weaker. A tax number, company registration, or email can belong to more than one legal entity (a parent and a subsidiary, two people at the same office). Those fields raise the score. They do not, by themselves, declare a match.

If both records have the same kind of ID in the same country and the values differ (two national ID numbers for “the same” name), Watchman lowers the score. The default multiplier is ID_CONFLICT_PENALTY_MULTIPLIER=0.70. Same name plus disagreeing passports is treated as two people.

A threshold you set as policy. minMatch is the lowest score Watchman returns. 0.80 is a high-precision screening line; about 0.59 returns more possible matches. You can document that choice in the risk assessment.

An audit trail. debug=true returns the pieces: which identifiers matched, what the name score was, whether an override or conflict penalty fired. That is the “effective challenge” artifact SR 11-7 asks for.

Open source. Apache 2.0. The scorer is Go you can read, test, and pin. You are not locked to a vendor’s unpublished matcher.

Evidence: 755,540 labeled pairs

OpenSanctions Pairs is the first large public pairwise benchmark on real sanctions/OSINT data: 755,540 expert-labeled pairs over 1,002,093 records from 293 sources in 45 jurisdictions, 76.9% positive, 64.7% cross-source, seven Unicode scripts (Smith, Sesodia, Lindenberg, Schroeder de Witt, 2026; dataset). Labels come from OpenSanctions analysts in production deduplication, not crowd workers.

We scored the entire dump with Watchman’s production Similarity (pairwise: if both records are already in front of the scorer, does the score agree with the analyst?). That is not a live OFAC-coverage test and not a blocking test. It is the cleanest public answer to “how does this matcher treat hard, labeled sanctions pairs?”

Subjects (472,477 people, companies, vessels — Occupancy/Succession auto-merge dropped) at threshold 0.80:

Precision is the share of returned hits that are real matches. Recall is the share of real matches that were returned.

Read that as a review-capacity vs miss-rate trade:

• Default Jaro–Winkler at 0.80: about 3,000 subject false positives and 94,000 missed labeled positives. Precision is exam-friendly; recall is not, especially on transliteration.
• Adding embeddings for different writing systems (qwen3-embedding:0.6b, used only when scripts differ): about 14,000 false positives and 56,000 misses. Precision stays 0.95. Recall on those pairs goes from half to 91%. Adding TF-IDF on top of that: about 4,400 more true matches and 3,400 more false hits; recall on different writing systems stays 0.91.

Name-algorithm swaps (Soundex, Double Metaphone, Beider-Morse, nsim, Editex) barely change precision or recall. They do not fix Arabic/Cyrillic/Latin pairs. That matches the paper: rule matchers over-fire on common Latin names; learned methods fail on transliteration unless you add a representation that is not character-based.

Vessels (7,550 pairs): precision 0.998, recall 0.825. IMO/MMSI still score 1.0 when both sides have them; pairs without those IDs go through name scoring.

Person/Person (284,808 pairs): precision 0.983, recall 0.683 at 0.80 without embeddings. That is where most remaining misses live, and where embeddings earn their keep.

Threshold as a documented control. Same Jaro–Winkler scorer, subjects only:

If the risk assessment says “we would rather review more alerts than miss a designation,” write down 0.59. If analyst capacity is the binding constraint, write down 0.80 and turn on cross-script embeddings.

How this sits next to the paper’s baselines

The paper reports sampled sets (about 1k–10k pairs) and includes auto-merge rows. Watchman figures here are the full 755,540-pair dump.

GPT-4o is a ceiling on pairwise identity when the model sees every field. Watchman is the list-driven control you can threshold, log, and defend. Wolfsberg still applies: screening compares names, IDs, and dates; it does not decide beneficial ownership.

Full tables: OpenSanctions Pairs evaluation and RESULTS.md.

What to tell an examiner

• Lists and refresh. Which files, which timestamp, how ingest of internal lists works.
• Fields used. Customer type, name and aliases, government IDs, DOB, address. Name-only is a degraded mode, not the design.
• Cutoff. minMatch and why (capacity vs miss-rate), with the OpenSanctions subject table as independent evidence.
• Transliteration. Whether embeddings are on, CROSS_SCRIPT_ONLY, and that hybrid recall on labeled cross-script pairs was 0.91 at 0.80.
• False-positive design. Tax IDs do not auto-confirm identity; conflicting national IDs penalize the score.
• Explainability. Stored debug pieces or equivalent logging of which fields drove the hit.
• Screening logs. Watchman does not retain search queries. You must keep who was screened, when, against which list hashes (GET /v2/listinfo), at which minMatch, and what was returned. That record lives in your application, not in Watchman.
• Change control. Scoring-policy changes are in git, tested, and described in methodology.

How to run it in the program

1. Screen onboarding, periodic refresh, and (where required) transactions against OFAC plus the other lists your risk assessment names.
2. Always send type= (person, business, vessel, …). It is optional in the API; omitting it searches every type, and a wrong type can miss the hit. Unknown type: call person and business (and vessel/aircraft when those IDs exist).
3. Send IDs and dates when CDD has them. That is the difference between a 1.0 passport hit and a name-only 0.81.
4. Set minMatch=0.80 with cross-script embeddings for a production-shaped queue; lower it if the board has accepted more review.
5. Route hits to investigation with the score pieces attached. True matches become blocks, rejects, and — where applicable — FinCEN reporting. False hits become tuning evidence.

Docker and config: Usage, Configuration. Search API: Search.

Sources

• Smith, Sesodia, Lindenberg, Schroeder de Witt. OpenSanctions Pairs: A Large-Scale Dataset for Pairwise Entity Matching. arXiv:2603.11051, 2026. Paper · Dataset
• Fellegi, I. P., and Sunter, A. B. “A Theory for Record Linkage.” JASA 64 (328), 1969. DOI
• Jaro, M. A. “Advances in Record-Linkage Methodology as Applied to Matching the 1985 Census of Tampa, Florida.” JASA 84 (406), 1989. DOI
• Winkler, W. E. String Comparator Metrics and Enhanced Decision Rules in the Fellegi-Sunter Model of Record Linkage. ASA Proceedings, 1990. ERIC ED325505
• FATF. International Standards on Combating Money Laundering and the Financing of Terrorism & Proliferation (Recommendations 6 and 7). fatf-gafi.org
• FATF. International Best Practices: Targeted Financial Sanctions Related to Terrorism and Terrorist Financing (Recommendation 6). Publication
• OFAC. A Framework for OFAC Compliance Commitments (May 2019). PDF
• OFAC. Economic Sanctions Enforcement Guidelines, 31 CFR Part 501 Appendix A. eCFR
• OFAC. FAQ 249, SDN search scoring. ofac.treasury.gov/faqs/249
• OCC / Federal Reserve. Supervisory Guidance on Model Risk Management (OCC Bulletin 2011-12 / SR 11-7). OCC
• The Wolfsberg Group. Guidance on Sanctions Screening (2019). wolfsberg-group.org
• FinCEN. Frequently Asked Questions Regarding Customer Due Diligence Requirements for Financial Institutions (2018). PDF
• FFIEC. BSA/AML Examination Manual — Risk Assessment. bsaaml.ffiec.gov
• FDIC. Bank Secrecy Act / Anti-Money Laundering resources. fdic.gov