Version 2026.08

Screening methodology

Names are matched against public sanctions, watchlist and PEP sources using an ensemble of name-similarity, phonetic, identifier and demographic signals, then enriched with public adverse-media, investigative-record and association data. Every finding cites its source.

Every report we produce cites the methodology version it was created under, and the text of a version never changes. If you need to defend a decision later, the rules that produced it are still here as they stood on the day.

Scope and status of this tool

  • TradeFlow OS is a decision-support tool, not a regulated screening provider and not a licensed commercial data vendor. It does not make compliance decisions for you and does not discharge any legal or regulatory obligation you hold.
  • Coverage is built from publicly available official sources. It is not equivalent to a licensed commercial dataset, which carries deeper politically exposed person (PEP) coverage, curated relative-and-close-associate links, and human-maintained profiles.
  • PEP coverage is derived from public structured data (Wikidata) and is materially thinner than a commercial PEP dataset, particularly for relatives and close associates. Treat a clean PEP result as the absence of a public record, not as confirmation that the subject is not a PEP.
  • A match is a similarity judgement about names and identifiers, never a confirmed identification of a person or company. Confirm every match against the cited source before acting on it.
  • A PASS records that no match was found in the sources listed, at the time shown. It is not a guarantee that the subject is unlisted, unsanctioned, or a suitable counterparty.
  • Adverse media, investigative records and association findings are research aids drawn from public reporting and leaked datasets. They are not findings of wrongdoing and carry no verdict.
  • Sources change constantly. A screen is a point-in-time result; only subjects under active monitoring are re-checked.

Data sources

  • Official sanctions, watchlist, export-control, debarment and forced-labour lists published by governments and multilateral bodies, refreshed on a schedule (daily for the fastest-moving US lists, weekly and monthly for the rest). A source that returns an empty or partial response is never allowed to overwrite good data, so a fetch failure cannot silently produce a clean result.
  • Politically exposed person data derived from public structured sources. See the scope limitations above.
  • Public news via the GDELT open news index, filtered to adverse themes and screened for source quality.
  • Investigative records from the ICIJ Offshore Leaks datasets. An appearance is an investigative signal, never a sanction.
  • Association links between listed parties and related persons, used to surface indirect exposure.
  • Where a paid Counterparty Due Diligence check is run, corporate registry, beneficial-ownership and extended AML data supplied by our verification provider.

Matching method

  • Candidate retrieval uses trigram similarity over a normalised index of every name variant, including aliases, so a query is compared against spellings the subject may not use themselves.
  • Normalisation transliterates Cyrillic, Arabic and Chinese scripts and strips diacritics and corporate suffixes, so a cross-script query still reaches the right candidates.
  • Candidates are then scored by an ensemble: whole-string and token-level name similarity, token-set overlap, edit distance, and phonetic similarity, combined with weights that favour token agreement over raw string distance.
  • Supplied identifiers adjust the score in both directions. Agreement corroborates a candidate; a conflicting value of the same identifier class (a different passport number, a different year of birth) reduces it, which is how two people with the same name are told apart. An agreeing IMO ship number is treated as near-conclusive because it is a globally unique, checksummed identifier that stays with a hull through renaming and reflagging.
  • Nationality and entity type are consistency signals: agreement gives a small lift, a clear conflict a small penalty, and a person-versus-organisation mismatch removes the candidate.

Verdicts

  • FLAG is returned at a confidence of 0.92 or above: a high-confidence match that must be resolved before proceeding.
  • REVIEW is returned from 0.80 up to 0.92: a plausible match requiring human judgement.
  • PASS is returned below 0.80: no match meeting the review threshold was found in the sources listed.
  • Thresholds are fixed and applied uniformly to every account, so a result means the same thing for every user and can be compared across time. They are not user-adjustable, because a screening threshold that each customer can lower is not a control.
  • Jurisdiction exposure is assessed separately from the list match and never changes the verdict. A subject can return PASS on every list and still sit in a sanctioned jurisdiction, and the report states both.
  • Where beneficial ownership is available, aggregate ownership by matched parties is evaluated against the 50% threshold used by OFAC. This is reported as exposure to be confirmed, never as a determination that an entity is blocked.

Decisions and record keeping

  • Every screen is recorded with its subject, verdict, matches, the sources searched, the time it ran, and the account that ran it.
  • A non-clear result is closed by an explicit decision: false positive, accepted risk, or confirmed. False positive and accepted risk require a written rationale, which is retained with the record.
  • Where a second-signature policy is enabled, a decision is not final until a second person approves it, and both identities are retained.
  • Reports are exportable as a fixed record showing the result and the methodology version under which it was produced.

Questions about a specific result

Every match we return names the list it came from, the date we last confirmed that list, and a link to the publisher, so any result can be checked at source. If a result looks wrong, that citation is the fastest way to establish it, and we would like to hear about it.

Contact support