Whitepapers
Sanctions Screening

Evidence, Not a Verdict

How on8's screening turns a name overlap into an auditable decision — and where it deliberately stops. A technical companion to "A Name Is Not Evidence", grounded in what the live platform actually does today.

on8.ai Team 01/09/2026Version 1.0

Get the paper

We'll email you a link to the PDF (2.0 MB). No account needed.

What's inside

  • A name overlap is a candidate, not a conclusion identity evidence decides the match
  • Three-valued evidence: "matches", "does not match", and "could not be compared" kept distinct
  • Coverage figures are read live from a running endpoint, not typed into a slide
  • Ongoing monitoring re-screens on a schedule and alerts only on a material change
  • An append-only audit trail, with every quantitative claim resolved to its live source
  • The limits named in the same voice as the claims the reason to trust the numbers

The wrong question, and the right one

Most sanctions-screening engines ask one thing: *how similar is this name to a name on the list?* The question that matters is different: *is this person the person on the list?* Our earlier paper, *A Name Is Not Evidence*, showed why the gap between those two questions is where compliance teams lose their hours — and where a tired team closes a real hit by the same reflex as the noise around it. This paper is the other half: how on8's screening is built against the opposite principle.

A candidate, not a conclusion

on8 treats a name overlap as a candidate, then tests it against identity evidence — date of birth, nationality, name coverage — and reports the result in three values, not two: *matches*, *does not match*, and *could not be compared*. That third value is the one most systems throw away, and it is the entire difference between a two-minute file and a twenty-minute one — between a closure rationale that survives an audit and one that does not. The engine's own response calls this "evidence, not a verdict": it produces a defensible basis, and a human closes the file.

Grounded in the live platform

Every headline figure in this paper is read live from a running endpoint, not a constant on a slide. The coverage numbers — 482 datasets and 167.1 million entities, measured 2026-08-13 — come straight from a count over the index, and the service refuses to publish a zero rather than show a stale figure. Around the engine sit ongoing monitoring that re-screens on a schedule and alerts only on a material change, configurable risk scoring, batch screening, and an append-only audit trail.

The limits, named on purpose

We wrote this the way we build: with the boundaries stated. Where a capability is fully realised, we point to the endpoint that serves it. Where it is partial, scoped, or deliberately switched off — death-based elimination surfaced to reviewers but not yet computed, the deep adverse-media tier switched off rather than faked — we say so in the same sentence, not a footnote. A screening vendor's honesty about what it does *not* do is the only reason to trust its numbers for what it does.

Download the full paper below.**

See it in your stack

Book a 15-minute walkthrough and we'll show you how on8 fits your compliance workflow.