Production Engine vNext · measured August 12, 2026
Performance, with the caveats attached.
The engine currently serving screening requests, measured through the production API path on real listed parties and verified adversarial near-misses.
Observed severe false positives
0
of 5,000
An observed result and an honest upper bound—not a guarantee.
300 / 300
Sanctions positives actionable
Every listed party in the tested sample surfaced
97.33%
Exact-name recall
95% CI 94.83–98.64%
0 / 150
PEP-only at sanctions grade
Political exposure remains distinct from sanctions
Decision distribution
Where 5,000 near-misses landed.
The adversarial names were designed to resemble listed parties but verified not to match one. Weak resemblance concentrated in the watch band; none crossed into a sanctions-grade result.
01
Clear
No action
02
Monitor
Low-confidence watch
03
Review
Analyst reviews evidence
04
Block
Sanctions-grade action
Before / after
Weak resemblance moved from a match to a watch.
Previous production engine
Near-miss names could surface at severe levels.
Resemblance carried too much weight, creating analyst noise and avoidable operational friction.
Production Engine vNext
0 / 5,000
severe false positives observed, while exact-name recall held at 97.33%.
Retrieval experiment · rejected
We built FAISS. The evidence said no.
Semantic vector retrieval was tested against the same corpus and turned off. It added a measurable precision cost without a recall improvement distinguishable from zero. vNext remains authoritative; FAISS remains disabled.
+4.4pp
Precision cost
More severe false positives
+0.1pp
Recall change
Confidence intervals overlap
Off
Production state
Rejected for this corpus
Methodology
How the claim was measured.
Real active listed rows formed the positive sample. Authored adversarial near-misses, verified against the current corpus, formed the negative sample. Both traveled through the same production-served screening path.
- 01
Prepare
Sample listed parties and verified non-matches.
- 02
Serve
Run every case through the deployed API path.
- 03
Classify
Map results into clear, watch and severe bands.
- 04
Bound
Publish counts with confidence intervals and limits.
Self-administered benchmark
Verifex designed and executed this benchmark on a corpus it authored. It is evidence about the measured cases, not independent certification or a promise of identical performance on every production query.
Limitations
What these numbers do not prove.
- 01
A zero-event result is bounded by its confidence interval. It is not proof that a severe false positive can never occur.
- 02
Results describe this measured corpus and deployed configuration. A different query mix or source distribution can move them.
- 03
The corpus has no identifier column, so this run measures name and context evidence rather than document-number corroboration.
- 04
This is a self-administered vendor benchmark and has not been independently audited.
- 05
Source availability is measured separately. An unavailable screening is never counted as clear or as a false positive.
Historical archiveMay 2026 Engine v3 synthetic benchmarkRetained for continuity. It is not the current production result.+
99.68%
F1
99.36%
Recall
100%
Precision
500
Synthetic cases
Test the engine on your own cases.
50 free screens each month. No credit card. Keep the evidence behind every result.