Evidence has an expiry date

We tend to treat a piece of evidence the way we treat a fact in a textbook: once it is true, it stays true, and we can reach for it whenever we need it. A balance is a balance. A reading is a reading. A signature is a signature. But almost none of the evidence that actually drives decisions is like a fact in a textbook. It is a description of a moment — a snapshot of a world that did not stop moving the instant the picture was taken. A credit balance describes an account at a particular second. A sensor reading describes a temperature, a pressure, a position that has already changed. A permission grant describes who was trusted under conditions that may no longer hold. A code-review approval describes a diff that may since have been amended. Each of these is true in the way a photograph is true: faithfully, and only of when it was taken.

This is easy to forget because the evidence does not announce its own age. It sits in a field, in a row, in a record, looking exactly as authoritative on the day it is consulted as it did on the day it was captured. Nothing about a stored value tells you whether the world it describes still exists. And so a decision can draw on evidence that was perfectly true, reason over it flawlessly, follow every rule it was bound by — and still land on the wrong answer, because the inputs no longer hold. The decision is procedurally immaculate and substantively false. That gap, between a decision that was made correctly and a decision that is still correct, is the whole of this essay.

The silent failure mode

What makes staleness dangerous is that it does not look like an error. When evidence is missing, a system can notice and complain. When evidence is malformed, a parser can reject it. But stale evidence is present, well-formed, and plausible. It passes every check a system knows how to run, because the only thing wrong with it is something the system was never asked to evaluate: whether it is still describing the present. A reading from an hour ago and a reading from a second ago are byte-for-byte indistinguishable as values. The difference lives entirely in time, and time is precisely the dimension that most systems flatten away the moment they store a fact.

So the failure arrives quietly. Nothing throws. No alarm sounds. A decision consults a value that was true and is no longer, treats it with the full confidence it would extend to a fresh one, and produces an outcome that is wrong in a way no one will detect until the consequences surface. The evidence was there. It was even accurate, once. It simply expired between the capture and the use, and no part of the pipeline was built to ask whether the milk had turned.

Stale evidence does not lie to you. It tells you the truth about a moment that has already passed, and lets you mistake it for now.

When was this captured

The remedy begins with a question that sounds like bookkeeping and is not: when was this captured? Provenance carries this as a plain fact — a capture-time, the moment the snapshot was taken. It is tempting to treat that timestamp as metadata, a clerical detail filed alongside the value, of interest mainly to auditors. But capture-time is not adjacent to the evidence. It is part of what the evidence means. A reading without its time is half a fact: it tells you what, and withholds the when that determines whether the what is still load-bearing. To know that a value was true is not enough to act on it. You have to know how long ago it was true, and whether that is short enough to still trust.

Once capture-time is in hand, freshness stops being a vague worry and becomes a question you can actually pose. How fast does this kind of fact decay? A geological measurement and a market price age at wildly different rates; an hour-old reading might be authoritative in one domain and worthless in another. Reasoning about that decay — naming, for each kind of evidence, how long it stays good — is not an optimization to bolt on later. It belongs to admissibility itself. The right question is not merely is this evidence accurate but is this evidence still admissible given how much time has passed since it was true. A decision that cannot answer that has not finished checking its inputs; it has only checked the ones that fail loudly.

Is it still right

This is where a replayable decision earns more than its usual keep. The familiar virtue of replay is that it lets you re-feed the same inputs through the same process and confirm the outcome follows — a way of asking was this decision right when it was made. But once capture-times travel with the evidence, replay can ask a sharper question. Not only "was it right," but "is it still right, given what we now know was already changing." You can hold the decision still and let the clock move: which of these inputs were near their expiry at the moment of decision, and which have since gone stale? A divergence on replay no longer just tells you the logic was wrong. It can tell you the logic was fine and the world moved — that the decision was sound on facts that had a shelf life no one had priced in.

That distinction matters because the two failures call for different repairs. A wrong rule wants a new rule. Evidence that expired wants a faster refresh, a tighter freshness bound, a decision deferred until the inputs can be re-captured. You cannot choose the right fix until you can tell the two apart, and you cannot tell them apart unless the record carried, all along, the time at which each fact was true.

A defensible decision, then, has to account for more than which evidence it used. It has to account for when that evidence was true, and whether that truth had any business surviving to the moment of use. Provenance that records capture-time, and a discipline that treats freshness as a condition of admissibility rather than an afterthought, are what let a decision answer the harder question honestly. Evidence has an expiry date whether or not anyone reads it. The only choice is whether the date is written down, reasoned about, and allowed to count — or quietly ignored until something built on a fact that had already turned comes apart, and no one can say why.

— Dispatches · Summit Cognitive