A model that was right once stops being right without announcing it

Every working model has a date on it

What people call intuition about a domain is mostly a compressed record of how that domain behaved while they were watching it. Someone who has run the same kind of project for a decade holds a set of expectations about timing, cost, failure modes and who to call. Those expectations were accurate, which is why they became automatic in the first place.

The trouble is that the accuracy was contingent. The model was fitted to a set of conditions — a market, a team, a technology, a set of rules, a body of habit — and none of those is stationary. When the conditions move, nothing in the model reports the change, because a model has no way of knowing that the world it was fitted to has been replaced.

Decay is slow and the signal is weak

A model that has gone out of date does not begin failing everywhere at once. It fails at the margins first, in cases that look like exceptions, and exceptions are exactly what an experienced person is comfortable absorbing. Every domain generates some noise, so a handful of surprises is consistent with a model that is still fundamentally sound. That tolerance is a feature and it is also the reason drift goes unnoticed.

The characteristic warning sign is not a single dramatic error. It is a change in the texture of the explanations you find yourself giving. When the exceptions start needing individual accounts, and when those accounts are getting longer, the exceptions have probably stopped being exceptions. A model that requires an escalating maintenance budget is telling you something about its fit.

This is not the same as refusing contrary evidence

It is tempting to file all of this under motivated resistance, but the mechanism is different and the difference matters for the remedy. In the familiar bias, someone avoids or discounts evidence against a belief they are attached to. Here there may be no attachment and no avoidance at all — the belief was simply correct, it earned its confidence honestly, and the evidence that would overturn it has not been especially prominent.

That is why exhortations to be open-minded do so little for this failure. The person is not closed. They are running a well-supported generalisation on inputs it was not fitted to, which is a calibration problem rather than a motivational one. The corrective has to be a scheduled re-examination of assumptions, because nothing in ordinary experience will raise the question spontaneously.

The opposite error is real and less discussed

The obvious response — revise fast, hold nothing firmly — produces its own failure, and a worse one in noisy environments. A model that is re-fitted after every surprise ends up tracking noise, which means it becomes less accurate while feeling more responsive. Anyone who has watched a forecast be adjusted weekly to whatever happened last week has seen this.

The judgement being asked for is therefore not “update more” but “distinguish drift from noise”, and the distinguishing feature is usually a mechanism. A surprise with a plausible structural story behind it — a rule changed, a supplier consolidated, a technology became cheap — is evidence of drift. A surprise with no such story is more likely to be variation, and variation deserves patience rather than a redesign.

Where the cost lands

The damage concentrates in exactly the places you would least like it to. Experienced judgement is trusted more, consulted more and questioned less, so a stale model has more influence than a novice’s uncertainty and is subject to less scrutiny. It also propagates: rules of thumb get taught, and a heuristic that has become wrong can outlive the person who formed it by years.

There is a common pattern in organisations where the informal rules describe conditions that ended some time ago, and nobody can say when they stopped being true because nobody wrote down what they depended on. The rule survives because it is cheap to follow and because its failures look like bad luck.

Making the assumptions visible enough to check

The practical move is to attach conditions to beliefs when you form them. Not merely “this supplier is reliable” but “this supplier has been reliable while their volumes were low and their key person was in place”. A belief stored with its conditions can be checked against the conditions, and checking conditions is a much smaller job than re-examining a belief from scratch.

The second move is to date things. A judgement formed several years ago about a fast-moving area deserves a different weight from one formed last month, and the age of a belief is information you already possess and almost never use. Ask when you last tested it. If the honest answer is that you cannot remember, that is the answer.