Methodology · reviewed 2026-11-15

How a name becomes a candidate — and why the score is similarity, not risk.

The same code runs at index time and at query time, so both sides of every comparison are normalised identically. Every candidate in a record carries the reasons produced by the steps below.

1. Normalisation

  • Unicode NFKD/NFKC folding, diacritics removed, upper-cased; “&” becomes AND; punctuation becomes spaces.
  • Legal-form synonyms fold to one token in both directions: CO ≡ COMPANY, LTD ≡ LIMITED, INC ≡ INCORPORATED, CORP ≡ CORPORATION, OJSC/PJSC/AO ≡ JSC, INDUSTRIES/INDUSTRIAL ≡ INDUSTRY, TECH/TECHNOLOGIES ≡ TECHNOLOGY and so on.
  • Chinese names produce variants: generic words translated (有限公司 → CO LTD, 硅业 → SILICON, 贸易 → TRADING), place names in their standard romanisation (新疆 → XINJIANG), the rest as pinyin with syllables paired the way companies usually romanise two-character words (合盛 → HESHENG). A USCC adds the registered name and former names from the mainland register.
  • UFLPA entries are split: aliases in parentheses become separate indexed names; “and subsidiaries” is kept as a flag.

2. Candidate retrieval

Postgres trigram similarity (pg_trgm ≥ 0.3) over the normalised names and aliases, OR any distinctive query token appearing as a whole word. Up to 600 names per variant are scored.

3. Scoring

StepRule
score_fullJaro-Winkler over the whole normalised string (p = 0.1, prefix ≤ 4), spaces removed.
score_partsEach query token is aligned to its best unused candidate token by Jaro-Winkler (pairs below 0.72 are not aligned); a Soundex-equal pair scores at least 0.9. Weighted average: identity tokens weight 1, legal forms 0.4, stop words 0.3; an unmatched query token keeps 0.8 of its weight with score 0; extra candidate tokens add 0.1 of their weight (legal forms 0.02).
name_scoremax(score_full, score_parts) × 100, rounded.
Country−10 when you give a country and the listed addresses name a different one; a note when they agree. No effect when the entry has no country.
IdentifierAn exact identifier hit (USCC, registration number in the entry's IDs) sets the score to 100.
TypeEntity / individual / vessel is a filter, not a score input.

Worked example: query “Hoshine Silicon” against “Hoshine Silicon Industry (Shanshan) Co., Ltd.” — tokens HOSHINE and SILICON match exactly (weight 2, score 2); the candidate's extra tokens INDUSTRY and SHANSHAN add 0.1 each, CO and LTD 0.02 each; 2 / 2.24 = 0.89 → 89. The record shows those reasons: “token 'HOSHINE' exact”, “token 'SILICON' exact”, “listed name also has 'INDUSTRY', 'SHANSHAN'”.

4. Threshold and verdict

Candidates at or above the minimum score (default 85, range 70–100) are shown and make the record's verdict POTENTIAL MATCH. None → NO MATCH FOUND, but only if every list was screened at a current or last-good version; if any source never loaded the verdict is INCOMPLETE. Near misses below the threshold are named in the method note. The record never says whether a transaction may proceed.

5. Three states per list

No match — screened at the version shown. Last good version — today's fetch failed (CSL/OFAC older than 48 h, UFLPA older than 7 days), the previous file and its date were used. Not screened — the source has never loaded. Inputs outside scope (EU/UK/UN lists, Hong Kong registrations, individual dates of birth) are reported as “Not covered”, not as no match.

6. Ownership look-through (Reviewed Record)

Shareholders come from the mainland register with their direct percentages. Each shareholder is screened. Under the OFAC 50 Percent Rule the direct holdings of listed (blocked) holders are summed; an entity reaching 50% joins the blocked set and its own holdings then count in full at the next level, iterated to a fixed point — never multiplied down the chain (FAQ 401 examples 1–5; FAQ 399: two 25% holders). Between 25% and 49.99%, or when any percentage is missing, the outcome is “review”, not “pass”. The BIS Affiliates Rule (15 CFR 744.11(a)(1)) uses the same arithmetic over Entity List and MEU holders but is stayed until 2027-01-10; results read “would fall under if the stay ends”. UFLPA is not an ownership rule: a listed shareholder is a review point, not a listing of the subject.

7. Record integrity

Every record is serialised with sorted keys and hashed with SHA-256 (the hash and exports fields blanked). Decisions you add rewrite the hash. The verify page recomputes the hash from the stored record.

What this is not

Not OFAC's algorithm: OFAC's name-parts technique is unpublished and its scores cannot be reproduced by third parties. Not an identity resolution: a candidate is a similar name until a person compares the entry. Not legal advice.

FAQ

Method questions

Is the score a risk score?
No. Score = name similarity between the query (or one of its variants) and the listed name, 0–100. A high score means the names look alike; whether the listed party and your supplier are the same entity is decided by comparing addresses, identifiers and context.
Why is the default minimum 85?
At 85 the common false alarms from shared generic tokens drop out while typos, legal-form differences and short-form names still surface. Lower it to 70 for a wide net; raise it to 95–100 to see only near-exact names.
How does the OFAC 50 Percent Rule example work?
FAQ 401, example 1: X (blocked) owns 50% of A, A owns 50% of B. A is blocked (50% by X). B is blocked because A — now a blocked person — owns 50% of B. The percentages are not multiplied (that would give 25%).

Handling false positives →