The jurisdiction of a number

A number is built somewhere. It comes out of a particular population, measured for a particular purpose, calibrated against a particular outcome that someone cared about for a particular reason. A risk score is not a fact about the world the way a temperature is a fact about a room; it is an instrument, ground to fit one job, and like any instrument it carries the shape of the bench it was made on. Inside that domain, the number means what its makers intended. Step outside, and it keeps the look of meaning while quietly losing the substance of it.

The trouble is that numbers do not stay where they are made. A score is portable in a way that its justification is not. The number is a single value, easy to copy, easy to paste into a new system, easy to treat as a property of the thing it was attached to rather than a property of the question that produced it. So it travels. A risk number built to decide one thing gets pulled into a decision about another, in a different context, against a different population, to answer a question its makers never posed. And it answers. It always answers, because a number does not know it has left home.

This is the move I want to name: a score built for one purpose is reused to decide another. It is rarely done in bad faith. It is done because the number is already there, already trusted, already paid for, and standing it up again from scratch for the new question would be slow and expensive. The existing number is right at hand, and it has a track record — in its own domain. The reuse feels like efficiency. It is, in fact, a quiet transfer of authority the number was never granted.

Validity has a domain

Every score has a jurisdiction, even when nobody writes it down. The jurisdiction is the set of conditions under which the number means what it claims to mean: the population it was built on, the outcome it was calibrated against, the time horizon it assumes, the purpose it was meant to serve, the base rates that held when it was made. Inside that boundary the score is doing legitimate work. Outside it, the score is an opinion wearing a uniform — and uniforms travel even better than numbers do.

The error is not in computing the number. The error is in asking it to rule somewhere it has no standing. A score built to rank one kind of risk, applied to a second kind, is not slightly less accurate; it is answering a question it was never validated to answer, and its confidence is borrowed from the old domain, not earned in the new one. The decimal places stay. The calibration does not travel with them. What you get is a number that looks exactly as precise as it did at home and is now precisely wrong, and there is no warning light, because precision and validity are different properties and only one of them is visible on the face of the score.

Authority is not a property a number carries everywhere it goes. It is granted by a domain, and it stops at that domain's edge — even when the number doesn't.

This is fundamentally a question of legitimacy, not accuracy. We tend to ask of a score, is it right? But a prior question governs it: is this the score's question to answer at all? Authorization to operate in one domain is not authorization to operate in every domain a number can physically reach. A number's legitimacy is local. It is conferred by the conditions that gave it meaning, and it does not renew itself automatically when those conditions change. A score that has wandered out of its jurisdiction may still be computable, still be defensible-sounding, still be acted upon — and have no legitimate authority over the decision it is now being used to make.

The record has to carry the jurisdiction

If a number's authority is bounded, then the boundary has to travel with the number. This is the practical demand. A score should never move through a system as a bare value. It should move as a value bound to its jurisdiction — what population it came from, what outcome it was calibrated against, what purpose it was built to serve, what assumptions it requires to hold. The Decision Receipt that records a decision should record not just which number was used but where that number was entitled to be used, so that the act of carrying it into a new domain is visible as the boundary-crossing it is.

This is what provenance is for. A number with provenance carries its own conditions of validity, which means a later reader — a reviewer, an auditor, the affected party — can check not only that the number was computed correctly but that it was applied legitimately. Provenance lets the boundary be enforced after the fact, by someone who arrives later and asks the right question: was this score being used inside its jurisdiction, or had it wandered out? Without provenance, that question cannot even be posed, because the number on the record looks identical whether it was at home or three domains away from anywhere it was ever validated.

And when the boundary is recorded, the reuse becomes contestable. That is the whole gain. The borrower, the patient, the counterparty acted on by a traveled number has a place to stand: not "the number was wrong" — which is hard to prove and easy to deny — but "the number had no jurisdiction here, and here is the record of where its authority ended." A boundary you can point to is a boundary you can defend. A boundary that lives only in the heads of the people who built the score, and dies the moment the score is exported, defends no one.

A number is not a fact you can carry anywhere and set down intact. It is a ruling, valid in one court, and a ruling does not bind a court that never granted it authority. Build the record so the number arrives with its jurisdiction attached — so that when someone asks it to decide a question outside its domain, the record answers before the number does: this is not yours to rule.

— Dispatches · Summit Cognitive