Only expand what an operator confirms: the rule that keeps an AI investigator on the right person

When people worry about AI investigators, they usually worry about the wrong thing. The fear is that the system will miss something. The real danger runs the other way: that it will find too much, about the wrong person, and present it with total confidence.

#The same-name problem

Every common name has a twin online. Search for a Jordan Lee and you'll find dozens of them, some in the same city, some with the same job title. A human analyst handles this with judgment: they notice the account is in the wrong country, or that the photo is clearly someone else.

An automated system has to handle it with rules, and a weak rule fails quietly. It picks the Jordan Lee whose profile looks most complete, and carries on.

#Why errors compound

In a single lookup, a wrong match is one wrong result. In an investigation, it's worse. Every finding becomes the starting point for the next step. If the engine wrongly decides that an account belongs to your subject, the next cycle searches that account's photos, its handle, its connections. Within a few cycles there's a detailed, well-sourced profile of someone who isn't your subject at all.

That's what we mean by building a second life for the wrong person. Every individual step looks reasonable. The result is wrong, and it might be shown to a client, an editor or an employer. (For how the loop moves from one step to the next, see Inside the loop.)

#The rule

Nuculair separates two kinds of steps.

Collecting around something already confirmed is automatic. If you've confirmed an email address belongs to your subject, the engine can freely check what that address is registered with.

Expanding the case to a new anchor waits for you. A reverse image match, a newly linked account, a handle that might be the same person: these are the moments where a case can jump to someone else. So the engine proposes them, shows its evidence, and holds until an operator confirms or rejects.

Your answer changes the loop immediately. Confirm, and the new anchor opens up new threads. Reject, and everything that would have followed from it is dropped.

#The checks that back it up

The rule doesn't work alone. Every candidate link also goes through a challenge step that argues the other side, looking specifically for namesakes, mismatched locations and timelines that don't fit. A link is accepted only above a confidence of 0.8 and rejected below 0.1, and every finding records how many independent sources agree rather than a bare "verified" stamp.

#Measuring how often we're wrong

A rule is a promise; a measurement is proof. We're building our evaluation around synthetic personas with deliberate same-name decoys, and the number we care most about is the decoy false-match rate: how often the engine attaches a decoy's details to the real subject. We'll publish it, from a recorded run, with the methodology and the misses.

#Where we draw the line

The same instinct shapes what Nuculair is for. We'd rather you start with your own footprint than with someone else's. Investigations are for professional work with a legitimate purpose, they're covered by our acceptable-use policy, and Nuculair can't be used for employment, credit, housing or insurance decisions. In Canada, the private sector's duties around personal information are set out in PIPEDA.

Correctness, consent and a human in the loop aren't features we bolted on. They're the reason the product works the way it does.