The half-life of a fact

We treat facts as if they were made of stone, but most of them are made of ice. A fact is a claim about the world that was true at the instant it was captured — the balance, the address, the status, the score, the reading — and the world it described does not hold still. Balances change, people move, statuses lapse, conditions reverse. The fact that was accurate when it entered the record begins, from that moment, to drift out of alignment with the reality it was supposed to represent. It has a half-life: a rate at which the probability that it still holds decays toward the probability that it does not. And a decision that rests on a fact without knowing how old that fact is, or how fast its kind of fact decays, is standing on a surface that may have melted while no one was watching.

This is easy to see with the obvious cases and easy to miss everywhere else. Everyone understands that an address from five years ago is probably wrong now, that a bank balance from last month tells you little about today, that a medical reading has a window beyond which it must be taken again. The half-lives there are short and intuitive. The danger lives in the facts whose decay is slow enough to feel permanent but real enough to matter — the qualification that was later revoked, the relationship that ended, the policy that was superseded, the identity that was reassigned. These sit in the record looking exactly as authoritative as the day they were true, and nothing in their presentation warns that their moment has passed. The fact does not announce its own expiry. It just quietly stops being true, and the record goes on presenting it as though it still is.

The failure, then, is not usually that a system used false information. It is that a system used information that was true once and is no longer, and could not tell the difference — because it never recorded when the fact was captured, never asked how long that kind of fact stays good, and never checked whether the window had closed before it acted. An input with no timestamp is not evidence; it is a rumor that happens to be filed. It asserts something about the world with no way to test whether that something has expired. And a decision built on a stack of such inputs inherits all their unmarked decay at once, arriving at a confident conclusion about a present that its evidence stopped describing some unknown time ago.

A fact with no date is a claim that refuses to say how old it is — and a decision that leans on it is trusting a witness who will not tell you when they last saw the thing they are testifying about.

Certainty that outlived its evidence

What makes stale evidence particularly treacherous is that decay does not diminish the record's air of confidence. A number that was measured precisely stays precise-looking long after it stops being current; the digits do not fade to signal their age. So a decision can be delivered with full apparent certainty on the strength of inputs that have quietly gone soft underneath it — and the certainty is not merely unwarranted, it is a kind of false statement. To present a conclusion as sure when its foundations have an unknown and possibly expired shelf life is to claim a confidence the evidence no longer supports. The precision is real; the currency is not; and the record, by omitting the second, lets the first stand in for both.

The remedy begins with a discipline that sounds almost too simple: every fact carries the time it was captured, and every kind of fact carries an estimate of how long it stays good. A Decision Receipt that records not just what an input said but when it was true, and flags an input whose freshness window has lapsed, tells the reader something the bare fact conceals — that this part of the foundation may have shifted since. That does not make the decision wrong. It makes the decision honest about the age of what it stood on, which is the precondition of anyone being able to judge whether it still holds. A conclusion resting on a fact three days past its half-life is a different object than the same conclusion resting on a fact captured this morning, and only a record that dates its evidence lets you tell them apart.

Provenance includes the clock

We already accept, for physical evidence, that provenance is not only about where a thing came from but about its condition over time — a sample has a chain of custody and also a use-by date, and a lab that ignored the second would have its findings thrown out however clean the first. The same standard belongs to the facts that feed a decision. Their provenance is incomplete if it records their origin but not their age, because origin tells you whether the fact was ever true and age tells you whether it still is, and a decision needs both. To know that an input came from a trustworthy source and nothing about when it was captured is to know half of what you need and to mistake it for the whole.

So the question to ask of any record is not only where did this fact come from but when was it true, and has that moment passed — because a system that cannot answer the second is asserting a certainty it may have lost without ever noticing the loss. Facts decay whether or not we mark their decay; the only choice is whether the record admits it. A record that dates its evidence and tracks its expiry can say honestly, "this rested on what was true as of then," and let a reader weigh the gap between then and now. A record that stays silent about the clock offers a conclusion frozen in a confidence its own inputs abandoned some time ago — and hopes no one asks how old the ice underneath it has become.

— Dispatches · Summit Cognitive