The decision that outlived its reason

There is a moment, in the life of nearly every rule, when the reason for it quietly dies and the rule goes on living. The threshold was set because a particular fraud pattern was surging in a particular quarter; the pattern subsided, the quarter passed, and the threshold stayed. The model was trained on a population that behaved a certain way; the population shifted, and the model kept scoring as though it had not. The exception was carved because a single vendor could not meet a single requirement; the vendor is long gone, and the exception is now load-bearing. In each case a decision that was once fully justified has become a decision that merely persists — not wrong, exactly, but no longer reasoned, because the reasoning that once held it up has expired without anyone marking the date.

We are careful, as institutions, about the birth of a policy. We convene the meeting, we document the rationale, we get the sign-off. What we almost never do is attach a condition of continued validity — a statement of the form this holds so long as the following remains true. And so the policy is born with a justification but no expiry, and it inherits the strange immortality of anything that is easier to keep than to revisit. The reason it was made is written down somewhere, in a memo no one rereads. The fact that the reason no longer applies is written down nowhere at all.

This is the difference between a decision and a standing decision. A decision is a thing you did once. A standing decision is a thing you are doing continuously, every time the rule fires, every time the model scores, every time the automated gate opens or closes. Each of those firings is a fresh act of authority — and it is being taken in the name of a reason that may have lapsed a year ago. The system does not know this. It cannot know this, because the only thing it retained was the rule, not the conditions under which the rule was warranted.

A rule that has outlived its reason is not neutral. It is an argument still being made on behalf of a premise that is no longer true.

Why decisions do not die on their own

The natural lifespan of a justification is short, and the natural lifespan of a decision is indefinite, and this mismatch is the whole problem. A justification lives in the world — in the conditions, the population, the threat, the constraint that made the decision the right one. Worlds change. But the decision, once encoded, lives in the system, and systems are built precisely so that things do not change unless someone changes them. The default is persistence. Inertia is not a bug here; it is the entire point of writing a rule down. What is missing is any mechanism by which the death of the reason propagates to the rule that depended on it.

Notice how asymmetric the effort is. Making the decision required a reason, a review, a deliberate act. Un-making it, when the reason expires, requires someone to notice the expiry, remember the decision, connect the two, and mount the case for change against the friction of an institution that treats every established rule as presumptively correct. Nobody is assigned to this. There is no role called keeper of expiry dates. So the noticing does not happen, and the decision accumulates a kind of unearned legitimacy simply by having survived — as though duration were evidence, when duration is only duration.

The most dangerous version of this is the decision whose original reason nobody even remembers. The threshold is 0.72 and no one alive knows why; it has been 0.72 for so long that it has stopped being a choice and started being a fact of nature. To question it feels almost superstitious, like moving a load-bearing wall in a house whose blueprints were lost. But a number no one can justify is not a fact. It is an orphaned decision, still deciding, in the name of a reason that has not merely expired but been forgotten — which is worse, because you cannot check whether a reason still holds when you no longer know what it was.

Provenance as the tracking of expiry

The remedy is not to make rules shorter-lived or to force a review of everything on a calendar, which only replaces reasoned expiry with arbitrary churn. The remedy is to bind each decision, at the moment it is made, to the reason that warrants it and to the conditions under which that reason holds — so that the justification is not a detached memo but a tracked dependency of the decision itself. This is what provenance is for, properly understood. It is not merely a record of where a decision came from; it is a live account of what the decision still rests on. A Decision Receipt that captures not only the conclusion but the premises makes expiry visible, because when a premise changes you can finally see which decisions were standing on it.

The point is to convert a silent lapse into an answerable question. When the population shifts, provenance lets you ask which decisions assumed the old population and get an answer, rather than discovering it one adverse outcome at a time. When the threat that justified a threshold subsides, provenance is what connects the subsiding to the threshold, so that the rule can be revisited on the merits instead of surviving on inertia. Expiry-tracking is not a feature bolted onto a decision; it is the difference between a decision you are still making for a reason and a decision you are merely still making.

So the question to ask of any long-standing rule is not has it worked — persistence guarantees it has appeared to work — but can we still state the reason, and is the reason still true. If the answer to either is no, the rule is no longer a decision. It is a habit with authority, an argument whose premise has quietly died while the conclusion goes on being enforced against people who never agreed to be governed by a reason that no longer exists. The kindest thing an institution can do is to let its decisions expire honestly, on the record, when their reasons do — rather than let them outlive the justification and pretend the survival was a verdict.

— Dispatches · Summit Cognitive