How EyesClear Investigations works

Version 1.0 · 1 October 2026

This describes the method, not the implementation. It does not reproduce our prompts or source code.

What a screening is

A screening answers one question about one named party: what does the public record say about this company or person?

The subject is chosen by our client — a regulated financial institution — because that party is, or is proposed to be, a customer or counterparty. We do not select subjects ourselves, and we do not select them by any personal characteristic.

Each screening is a live lookup. We do not maintain a database of people.Nothing is pre-computed, and no profile accumulates between screenings.

What it searches

Reference lists, held by us and refreshed daily from the issuing authority: the OFAC Specially Designated Nationals and consolidated non-SDN lists, the UN Security Council Consolidated List, the EU consolidated financial sanctions list and Annex IV to Regulation 833/2014, the UK OFSI consolidated list, and US export-control lists from the Bureau of Industry and Security and the State Department.

Registers and archives, queried live by name: company registers, insolvency and bankruptcy notices, court archives, watchlist aggregators and published news.

The open web, searched by a language model with a web-search tool, which reads the pages it finds and cites the passage each statement rests on.

A screening does not run every source. Which sources are worth consulting depends on the subject — its type, where it operates, and what the client asked — and the report records which ran, which were skipped, and why.

Where a model makes a judgement, and where it does not

We separate the mechanical from the evaluative, deliberately.

Mechanical, done in code: searching a name; retrieving records; filtering sources by whether they can apply to this kind of subject at all. These steps are designed to be generous — they surface candidates rather than settle them, because a filter tuned to reject look-alikes will also reject real matches.

Evaluative, done by a language model, and disclosed as such:

• Identity. Does this record concern the subject, or a different party with a similar name? This is the single most consequential judgement in the system.
• Seriousness. What stage did the matter reach, as the record states it — an allegation, an investigation, a charge, a penalty, a conviction?
• Relevance. Does the source say something a compliance analyst needs, or does it name the subject only in passing?

We use commercial language models from established providers for these steps. No model is trained on our clients' queries or on the reports we produce.

Never done by a model: predicting the outcome of any case, advising on any legal position, or deciding anything about a person. Those are outside the service by design.

What happens to a candidate record

1. A search returns candidates for the subject's name.
2. Each candidate is assessed for identity. Those judged to concern a different party are set aside and recorded — they appear in the run detail as considered-and-rejected, and never in the report as findings.
3. Surviving matters are placed on the severity scale and carry the source, a link to it, and a quotation supporting the point reported.
4. Where identity cannot be settled, the matter is kept and labelled unconfirmed, requiring review. It is never silently discarded and never presented as established.
5. The report is assembled with the coverage bounds for every check that ran.
6. A human analyst reads it and decides what to do. The service does not act on anything.

The severity scale

Findings are placed on a scale that describes what the record shows, not our view of the person:

• Unverified claim — an assertion we could not tie to an authority's own record.
• Allegation — a claim made in proceedings or reporting, not determined.
• Investigation — an authority is or was examining the matter.
• Charge — a formal accusation has been brought.
• Penalty — a sanction, fine or regulatory measure has been imposed.
• Conviction — a court has determined the matter.

An allegation is not a finding of wrongdoing. Nothing in a report asserts that any person is guilty of anything.

What a report cannot tell you

Absence is not evidence of absence. A screening that returns nothing means we found nothing, not that nothing exists. Sources go missing, sit behind paywalls, publish in languages or registers we did not reach, or are simply wrong.

Coverage is bounded, and every check says how. For example, UK court judgments are searched through the National Archives' Find Case Law service, which covers England and Wales (plus UK-wide Supreme Court and Privy Council decisions), mostly from the early 2000s; Crown Court and County Court judgments are received only for selected cases and magistrates' courts are not routinely included; many decisions are given verbally and never transcribed; and settled or withdrawn proceedings produce no judgment at all. A clean court result is not a criminal record check. Every other source carries its own bound in the same way.

Identity groupings are an assessment, not a certainty. Names are shared. Where we group findings under one identity, that is our system's best assessment and is stated as such.

A report is a snapshot. It reflects what was findable at the moment it ran. Sources change, and we do not monitor them afterwards or update reports already produced. Because we hold no copy of any source collection, a record withdrawn or anonymised at source simply stops being returned to us.

Sources are linked, not replaced

Every finding links to the source it came from. We do not reproduce records in full, host copies, or index their text. Quotation is limited to a short extract supporting the specific point reported.

The report tells the reader to check the cited source before acting on anything. That instruction is the point of the design, not a disclaimer attached to it.

Testing and review

We maintain a fixed set of test cases with known-correct outcomes and re-run them whenever the models or the method change, so that a change in behaviour is detected rather than assumed.

We will test specifically for unequal treatment across communities, by re-running those cases with subject names substituted across linguistic and national origins and comparing the outcomes.

Our use of court records is reviewed annually against the Ministry of Justice's nine principles for computational analysis, and the outcome of that review is recorded. Where it changes how the service works, this statement is updated to match.

Licences and attribution

We hold the licences our sources require and use each source within its terms. Where a licence requires attribution, we give it:

• Contains public sector information licensed under the Open Government Licence v3.0.
• Contains information licensed under the Open Justice – Licence v2.0.

If a report is wrong about you

If you appear in a report and believe we have misread a source or matched the wrong person, write to informationsecurity@eyesclear.comwith the report reference and we will investigate. Where the fault is in our reading or our matching, we correct it. Where the fault is in the source, correcting the record usually belongs with the source itself, and we will say so. Nothing here limits any rights you have under data protection law.

Further detail is available to The National Archives and to regulators on request.