01
Normalize
Both the query and the list entry are reduced to a comparable form: transliteration to Latin script, Unicode NFD normalization, diacritic removal, punctuation stripping and whitespace collapsing.
A screening result is only defensible if the method behind it can be explained. This page documents the normalization, candidate generation and scoring MatchAudit applies, what a confidence value does and does not mean, and the limitations that keep a human reviewer accountable for the decision.
Matching pipeline
Each stage is deterministic and inspectable. The output of the pipeline is a candidate list with rationale, not a verdict.
Both the query and the list entry are reduced to a comparable form: transliteration to Latin script, Unicode NFD normalization, diacritic removal, punctuation stripping and whitespace collapsing.
A fuzzy index built over primary names and published aliases returns candidates above a strict similarity threshold. A token prefilter limits comparisons to entries that share meaningful name tokens.
Each candidate carries a confidence value derived from the similarity score, a label, the matched text fragments and a human-readable rationale for why it surfaced.
Your reviewer compares identifiers, records an outcome and keeps the rationale. MatchAudit does not decide the outcome and does not adjudicate on your behalf.
Configuration
These are the choices a reviewer needs in order to judge whether the result set is fit for their risk appetite. Each one is a trade-off, stated with its cost.
Candidate generation uses Fuse.js fuzzy search configured conservatively — a 0.10 threshold on the primary screening path and 0.12 on the unified screening service, with a bounded edit distance and a two-character minimum match length. Tight thresholds mean fewer weak candidates to triage, at the cost of rejecting very distant spellings.
A name of three or more tokens must corroborate on at least three tokens, unless a first-name and last-name pair already corroborates. This suppresses matches driven by a single common name component.
No Soundex, Metaphone or comparable phonetic algorithm is applied. This keeps rationales explainable and avoids the over-broad hit volume phonetic keys produce, and it means phonetically similar but textually distant spellings can be missed.
Managed sanctions sources are refreshed on scheduled jobs — most several times per day, some once daily. If a refresh fails, screening continues against the last successful snapshot and the dataset state stays visible rather than failing silently.
Same registry on every plan. Free, Starter, Business and Enterprise screen against the same managed source registry. Paid plans add API access, higher-throughput workflows and evidence features — not a different or larger dataset.
Scoring
Confidence is derived from the fuzzy similarity score as one minus that score, then labelled. It describes how closely two strings agree after normalization — nothing more.
Confidence
0.95 – 1.00
Very High
Near-exact agreement on the normalized name. Still requires identifier review before any decision.
Confidence
0.85 – 0.949
Medium
Substantial but incomplete agreement. Typically where date of birth, nationality and programme comparison does the real work.
Confidence
below 0.85
Low
Weak textual agreement. Surfaced for completeness, not as an indication that the subject is listed.
Where a publishing body includes them, these fields are parsed and attached to the candidate so a reviewer can rule a match in or out on evidence rather than on the name alone. Availability differs by source; the live inventory shows what each dataset carries.
Dataset provenance, source timestamps, the matched text fragments and the rationale travel with the result into the audit trail and the Evidence Pack, so a later reader can see which dataset state produced the candidate.
Review a sample Evidence PackStated limitations
A matching engine that claims no limitations cannot be assessed. These are the boundaries a reviewer should design their procedure around.
See the full assurance boundaryMatchAudit produces candidates, rationale and a traceable record. Your team compares the identifiers above, records an outcome against your own policy and owns that decision. MatchAudit does not staff reviewers, does not adjudicate matches and does not provide legal or regulatory advice.
Related disclosures
Data coverage, security controls, privacy terms and contract language are maintained as separate records so a reviewer can check each one independently.
Live inventory
The authoritative source list with record counts and per-source last-update timestamps.
Open resourceJurisdictions
Which publishing bodies and jurisdictions are covered, and how current each dataset is.
Open resourceSecurity
Access control, storage, file intake, AI boundaries and the claims we deliberately do not make.
Open resourceOperations
Recoverability, incident handling and the boundary between operating objectives and contractual commitments.
Open resourcePersonal data
Controller and processor roles, purposes, legal bases, retention and data-subject rights.
Open resourceContract
The agreement governing use of screening outputs and the allocation of decision responsibility.
Open resourceAssessing this method for a procurement or model-risk review? Tell us the population you screen, the volume and the evidence your reviewers need, and we will answer against the current implementation rather than a blanket claim.
Last reviewed September 22, 2026. This page describes the current MatchAudit implementation and is updated as thresholds, sources and scoring evolve. It is informational and does not vary the Terms of Use.