A false positive is a hit on someone who is not the listed person. They cost reviewer time and, worse, train teams to click "dismiss" without looking. The fix is not a stricter engine; it is better inputs and honest scoring.
Where they come from
- Common names. Lists contain hundreds of people named Mohammed, Kim or Ali. A name-only query on Kim Jong Un returns ten KIMs above 0.7 in our data, because that is what the lists hold.
- Missing supporting facts. Without a date of birth or nationality the engine cannot separate two people with the same name.
- Weak aliases. Publishers mark some aliases as weak (nicknames, partial names). Treating them like primary names inflates hits.
- Thresholds set for comfort. A 0.5 threshold feels safe and doubles the review queue.
What actually reduces them
- Send the date of birth. In KYCWire a conflicting birth year applies a penalty, so a same-name candidate born twenty years apart drops out of the likely band. This is the single largest lever.
- Send the country. Same mechanism, smaller effect.
- Send identifiers when you have them. A passport or registration number hit ranks first and lets you dismiss name-only candidates with confidence.
- Use two bands. Auto-clear below 0.7, auto-escalate at or above 0.85, human review in between. Measure how often the middle band turns out real and move the lines.
- Specify the entity type. A query typed
Organizationis never scored against persons.
What not to do
Do not raise the threshold until the queue is quiet. False negatives are invisible until a regulator finds one. Our public test set keeps recall at 0.99 at the default threshold; we publish the misses on the methodology page so you can judge the trade-off with data.
