Drift is a silent edit

We tend to think of drift as a slow degradation, a kind of weathering. The model gets a little less accurate as the world moves away from its training distribution; the policy ages; the data source updates its schema. We reach, almost reflexively, for the language of decay — drift as rust, as erosion, as something fading on its own. But that vocabulary hides what is actually happening, and the hidden part is the part that matters. When the model, the policy, or the source underneath a past decision changes, the past decision does not fade. It changes. Someone, or something, has revised the basis on which it was made — after the fact, without notice, and without a name attached to the revision.

Consider what it means to look back at a decision once the ground beneath it has moved. You ask: was this right? To answer, you reach for the rule it followed, the model it consulted, the data it was given. But the rule you reach for is the current rule. The model is the current model. The source is whatever the source says today. You are not examining the decision. You are examining a reconstruction of the decision built out of present-day parts — and you cannot tell the difference, because nothing in the record told you the parts had been swapped. The decision now appears to have been made on grounds it was never made on. It has been edited. And the edit has no author, no timestamp, no record of itself.

Decay forgives; an edit does not

The reason the distinction matters is that we forgive decay and we do not forgive edits. If a record simply got fuzzier with age, we would discount it accordingly — old evidence, handle with care. That is a manageable problem. But an unauthored edit is a different kind of injury. It does not make the record less certain; it makes the record assert something false with full confidence. The decision now reads as if it followed today's policy, scored against today's model, and weighed today's data, when it did none of those things. Drift does not blur the picture. It paints over part of it and signs nothing.

A record that does not freeze its world does not preserve a decision; it preserves a sentence whose meaning anyone can change later.

This is why provenance cannot only mean knowing where something came from. It has to mean holding that origin still. A source citation that resolves to "wherever this lives now" is not provenance; it is a forwarding address that points at a moving target. By the time you follow it, the thing it pointed to may have been updated, corrected, re-weighted, or withdrawn — and the citation will report all of those silently, as if they had always been true. The pointer keeps its shape while the world it points into is replaced underneath it. Provenance that does not freeze is provenance that lies on a delay.

The only defense is to hold the world still

So the defense against drift is not faster monitoring or more frequent retraining, useful as those are for the future. Monitoring tells you the present has moved. It does nothing for the record of the past, which is already being rewritten by the same movement. The only defense for a past decision is to have frozen, inside the record itself, the world that decision was made in — the actual evidence as it stood, the rule in the exact version that bound it, the model's verdict as it was returned, enough of the surrounding state that the decision can be replayed against its own moment rather than against ours.

This is the quiet discipline behind a Decision Receipt that is built to last. It does not point outward at sources that will move; it carries its world inward, sealed at the instant of decision, so that the basis cannot be swapped out from under it. When you later ask whether the decision was right, you are handed the grounds it actually stood on, not a reconstruction assembled from present-day parts. The receipt does not protect you from the world changing. The world will change; that is not a thing a record can stop. It protects you from the world changing silently — from the revision happening without anyone deciding to make it, naming it, or being answerable for it.

And there is a sharper point underneath the practical one. An edit nobody authored is an edit nobody can be held to account for. If a person rewrote the basis of a past decision, you would want their signature on the change, and you would ask why. Drift performs exactly that rewrite and presents no one to ask. The change has no author precisely because it was never decided — it accumulated, one schema migration and one model update at a time, until the past quietly said something it had never said. The absence of an author is not innocence. It is the thing that makes the edit impossible to contest, because there is no claim on the table to dispute, only a present that calmly insists it was always the past.

So when you go to defend a decision and find the rule has changed, the model has changed, the source reads differently than you remember — do not file it under things getting old. Recognize it for what it is. Something has edited your history and left no note. The only records worth trusting over time are the ones that made that edit impossible by refusing to depend on a world that would not hold still. Freeze the world a decision was made in, or accept that the decision will be rewritten by whoever changes the parts — which is to say, by no one, on no one's authority, with no one to answer for it.

— Dispatches · Summit Cognitive