A decision is not a prediction

A weather forecast that says there is a seventy percent chance of rain has not decided anything. It has not told you to carry an umbrella, cancel the picnic, or delay the flight. It has reported a probability, and then it has stopped, leaving the actual choice — to act, and how — entirely to you. This is the honest condition of a prediction: it describes the world and declines to commit to anything in it. The committing is a separate act, performed by someone else, and the gap between the two is where all the responsibility lives. We are losing sight of that gap, and the loss is not innocent.

A model, in the technical sense, produces predictions. It takes inputs and returns a distribution, a score, a likelihood — a statement about what is probably the case or what will probably happen. That output, however sophisticated, has the same logical character as the weather forecast. It is a description that stops short of commitment. The model does not deny the claim, approve the loan, flag the transaction, or remove the post. It emits a number, and the number, by itself, does nothing to anyone. Whatever happens next is the work of an institution that takes the number and acts.

The act is a different object than the output

The decision is that act. It is the moment the probability is converted into a consequence — the moment someone or something draws the line that says this score is high enough to deny, this likelihood is sufficient to flag, this distribution warrants removal. That conversion is not contained in the model's output. It is a judgment laid on top of the output: a threshold chosen, a policy applied, a commitment made on the strength of a number that, on its own, committed to nothing. The prediction and the decision are different objects, and the difference is not academic. One describes; the other does. Only the one that does has a victim, and only the one that does is owed an account.

This is precisely the distinction that gets quietly collapsed when we say a system "decided" something, as though the deciding were a property of the model. It is not. The model predicted. The institution decided — by building the system to act on the prediction in a particular way, at a particular threshold, with particular consequences attached. To attribute the decision to the model is to misplace the agency, and the misplacement is convenient in a way we should distrust, because it relocates responsibility from the party that made the commitment to a thing that, properly understood, made no commitment at all.

A probability is owed no apology and answers to no one. The instant it is turned into an act with a consequence, that immunity ends — and it ends for the party who turned it, not for the number.

Probabilism is not a shield

The reason this matters so much in practice is that the language of prediction carries an exculpatory tone that the language of decision does not. A prediction is humble by construction; it speaks in odds, it admits uncertainty, it never quite says yes or no. So when a consequential decision is dressed in the model's probabilism — "the system assessed a high likelihood," "the score exceeded the threshold" — it inherits that humility, and the humility functions as a defense. The decision gets to hide behind the prediction's frank acknowledgment of uncertainty, as if the uncertainty diluted the responsibility. But it does not. The person whose claim was denied did not receive a probability. They received a denial. The act was definite even though the prediction was not, and the definiteness of the act is what creates the obligation to account for it.

This is why the standard owed to a decision is categorically different from the standard owed to a prediction. We do not ask a forecast to justify itself; we ask only whether it was well-calibrated over many trials, and a forecast that is right seventy percent of the time when it says seventy percent has met its entire obligation. But a decision cannot answer to calibration alone. A decision answers to the specific person it landed on, in the specific instance, and the question it must be able to face is not "are you usually right" but "why this, here, now, to me." That question is not addressed to the model. It is addressed to whoever turned the model's number into the act — and they cannot answer it by pointing back at the number.

So the record has to capture the conversion, not just the prediction. It is not enough to log that the model produced a score; the account has to show the moment the score became a decision — the threshold that was active, the rule that bound the score to a consequence, the evidence the score itself was computed from, all preserved as they stood. This is what a Decision Receipt is built to hold: not the model's opinion, which commits to nothing, but the institution's act, which committed to everything. The prediction is upstream and the institution may borrow it, cite it, lean on it. But the decision is its own, and the account it owes belongs to it alone — because the affected party was harmed by the act, and the act is the only thing that can be asked to answer for itself.

— Dispatches · Summit Cognitive