Methodology
One method, the same for everyone.
A test result is only worth what its method is worth. This page sets out how test sets are built, run and scored, how we stop anyone tuning to the test, and who sees what.
Methodology v1.1 · updated 28 October 2026
What each baseline covers
Sanctions
33,000+ records, 4 lists
PEPs
c. 3,000 hand-checked, growing to c. 7,500
Name matching
c. 75,000 variants and non-matches
Adverse media
c. 2,000 hand-checked subjects, from Q1 2027
Agentic performance
Seeded alert sets, from Q1 2027
Baseline, then a retest every quarter against anonymised peers.
01 · Where a miss happens
Six places a result can go wrong, each with a different fix.
A person your screening should have caught doesn't alert. Every miss is traced to the stage where it failed, which tells you who fixes it.
Blind
You receive test IDs, names and the fields your system accepts. Expected results never leave us.
Verified at the source
Every person in a test is checked against an official list, register or gazette before it enters a test.
The right person
A hit counts only when it's the right person. Finding a namesake isn't finding them.
Every miss has an owner
The data, the engine, the configuration or, from Q1 2027, the AI agent. Never one blended number.
02 · Two kinds of test
Matching and coverage are scored separately.
Matching
Sanctions and name matching
The person is certainly on the list. The question is whether your engine finds them when the name arrives in another form, and what that costs in alerts on ordinary names. A miss here is an engine or configuration fix.
Coverage
PEPs, relatives and close associates, adverse media
The question is whether the person is in your data at all, as the right person, with the right details. A miss here is usually a data fix.
03 · Building test sets
Real names, and variants that earn their place.
Sanctions. Drawn from the UK Sanctions List, OFAC SDN, EU consolidated and UN consolidated lists: 33,000+ designations covering c. 25,000 people, companies, vessels and aircraft. That includes c. 105,000 published names and aliases (c. 14,700 of them in original script), c. 2,300 vessels by IMO number and 342 aircraft. Every record is checked against the official publication, never against another provider's data. A record counts only when it's the right designation with the right identifiers.
List counts as at 27 October 2026.
Name matching and false positives. A baseline built to c. 50,000 name variants of real listed people and c. 25,000 ordinary names drawn from real customer geographies. Every variant has to justify its place: where it arises in real data, which naming populations it affects, and what result is expected before the run. Results are broken down by naming tradition, including Chinese, Korean, Russian and Ukrainian, Arabic, Persian and South Asian.
PEPs. c. 3,000 hand-checked PEPs at baseline, growing to c. 7,500. Real office-holders verified against parliaments, government gazettes, courts and company registers, not aggregated sources such as Wikidata. Relatives and close associates are rarely published officially, so they're verified against reputable news reports naming both people. Covers the US, the UK, Canada and all 27 EU member states, with Saudi Arabia and the UAE from 2027 and other countries on request. The roles that become PEPs under the EU AMLR from 10 July 2027 are scored separately. Names found only by a live web lookup are reported separately, not in the headline.
Adverse media. From Q1 2027. About 2,000 hand-checked subjects across fraud, bribery and corruption, money laundering, sanctions evasion, tax and organised crime, by country and language. Each is paired with near-namesakes and with people named in the same story who aren't its subject, such as victims. It measures coverage, whether the story is attached to the right person, and how quickly new stories appear.
Agentic performance. From Q1 2027. Seeded alert sets, with true matches hidden among false positives, test whether an AI agent reviewing alerts escalates the real matches, clears the rest, decides the same way each time and cites the right evidence. Hard cases include near-namesakes, partial identifiers, transliterated names and conflicting dates of birth. A true match the agent clears is reported first, because it's the costliest error.
04 · Running the test
One blind file, run at your normal settings.
Mixed batches
Test names are mixed into larger batches so no single name stands out.
Control names
People who share or nearly share a listed name, to check over-matching.
Canary names
Invented people. If one ever appears in a data set, the test set has leaked.
Your settings, recorded
We record the lists, threshold and matching settings you ran with, and the date. The expected result for each name is fixed for that setup on that day.
05 · Scoring
Checked controls first, then the findings.
- Exact-name and blank controls must pass before any result is reported. A failed control makes every other number uninterpretable.
- Matching and coverage are scored separately, and never combined into one number.
- A match needs the name plus role, country or date of birth. Cases where the family-name position can't be established are skipped, not guessed.
- A namesake doesn't count as finding the person. Where the test supplies a date of birth or country, an alert on a namesake also counts as a false alert.
- Each miss is placed at one of the six stages in section 01.
- False alerts are reported alongside misses, using control names and ordinary names, as alerts per 1,000 names.
- Results found only by a live web search are reported separately from records held.
- A category or country result appears only with at least five names behind it, and is always shown as a count (for example "4 of 5") next to any percentage.
- Peer comparisons appear only once at least three other providers have been tested. Peers are shown unlabelled, and comparisons describe the providers and sample tested, not the whole market.
- Every result is dated, so it can be quoted.
06 · Keeping it fair
Improve your data. Don't tune to the test.
About half
Fixed core, same every cycle
The rest
New or rotated each cycle
Sanctions name variants rotate too. A held-back set, whose names are never disclosed, runs alongside every cycle.
Everyone tested agrees not to make changes just to pass, such as adding only the names tested. Results are checked for signs of targeted fixes: gains on the core that don't show up on the held-back set are flagged or excluded, so a published result stays comparable with everyone else's.
07 · Independence
Paid for reports, never for scores.
Firms and providers pay for their own reports and retests. Every one is scored by the same method, whether they take part in a programme or not, and no one can buy a different result or a say over anyone else's.
Results are never shared by name without the owner's agreement, and no provider ever sees another provider's results by name. Anonymised, aggregate findings may be published. A named public claim carries the test date and a link to this page, and its wording is agreed in advance. Individual test names are never published.
We use AI tools to help analyse results only where you've agreed to it. Every published number is checked by hand against the underlying cases.
Questions
What people ask before a first test.
Which banks and firms see a provider's results?
Only those the provider chooses to show, by publishing or sharing its report. Firms that test their own screening see results for their own setup. Everyone else sees anonymised, aggregate findings.
Do you need our customer data?
No. Our test names go through your screening, or a test API key you provide. No customer data leaves your systems.
Can a vendor or provider tune to the test?
Not without it showing. About half of each PEP and adverse media set is new or rotated every cycle, sanctions name variants rotate, a held-back set is never disclosed, and gains that appear only on known names are flagged.
How is it priced?
The quarterly testing & benchmarking programme is an annual subscription priced by scope: data, markets and categories.