Current production methodology
Match candidates. Assess identity. Preserve the decision.
Verifex separates retrieval from identity assessment, and identity assessment from the customer-facing decision. A useful screening result must also state what actually ran and what its coverage permits.
Current production configuration
Production Engine vNext uses lexical + fuzzy + phonetic retrieval. FAISS semantic retrieval is disabled in the current production configuration.
Older engine labels can remain in historical records. They are historical context, not a description of the current production method.
The decision path
A candidate is only the beginning.
- 01
Retrieve candidates
Current production retrieval combines lexical, fuzzy and phonetic paths. It can retrieve a candidate for review, but retrieval alone does not establish identity or sanctions status.
- 02
Assess identity evidence
Name similarity is evaluated with available context such as aliases, date of birth, country, nationality, entity type and identifiers. Missing evidence remains missing. A conflicting identity field cannot silently confirm a candidate.
- 03
Apply the canonical decision
The served decision considers execution, coverage, candidate and identity state, policy and evidence availability. A legacy score or local risk tier cannot weaken a stricter served verdict.
- 04
Preserve what happened
The result records source execution and coverage alongside the decision. Evidence supports review and reconstruction, it does not turn a similarity score into a legal conclusion.
Decision discipline
What the result must never imply.
Candidate is not confirmation
A retrieved name is a lead for identity assessment. It is not a confirmed identity simply because it has high similarity.
No candidate is not always clear
A no-candidate result cannot establish clearance when coverage, execution, candidate-set completeness or mode limitations are incomplete or unknown.
PEP remains a separate risk category
PEP evidence is not sanctions evidence and is never presented as wrongdoing, a sanctions designation or an automatic block.
Measured current path
Production benchmark, with limits stated.
Measured August 12, 2026 on the deployed engine: 0 severe false positives across 5,000 adversarial negatives, 97.33% exact-name recall. The corpus covered 1.3M+ active records across 50+ sources.
0/ 5,000
Severe false positives across adversarial negatives (95% CI ≤ 0.077%)
97.33%recall
Exact-name recall held on the production path
0/ 150
PEP-only cases served at sanctions grade
Measured August 12, 2026 through the production-served API path. Severe false positives are bounded by their 95% confidence interval, not claimed impossible. View the benchmark.
Known limits
Coverage and uncertainty stay visible.
The benchmark is vendor-authored and self-administered. It is not an independent audit or a guarantee of future performance.
Coverage is result-specific. A source that is unavailable, excluded, partial or unknown cannot be represented as clean or complete.
Exact-Only mode is a constrained retrieval mode. An exact-only no-hit is not a general compliance clearance.
Script and transliteration behaviour is only described publicly when it has been measured end to end. This page does not make a blanket cross-script coverage claim.