Record the alternatives you rejected

Look at what your decision system writes down when it acts, and you will almost always find a single value. Approved. Denied. This route, this action, this account, this tier. The one selected result, recorded cleanly, and everything else the system had in front of it at the moment of choice discarded the instant the choice was made. This is the natural shape of a log — a log records what happened, and what happened is that one option won. But a decision is not only the option that won. It is that option chosen over specific others for specific reasons, and a record that keeps only the winner has thrown away the part that made it a decision at all rather than a foregone conclusion wearing a decision's clothes.

The directive is this: capture the alternatives the system considered and rejected — the other options that were on the table, and the scores or reasons that put them below the one you kept — not merely the output that prevailed. It is a build decision, and like most of them it is made by default, invisibly, in the wrong direction. The comparison happens inside your code; the ranking is computed, the branches are weighed, the losing options are held in memory for exactly as long as it takes to pick the best one. Then they fall out of scope and are gone. You had the whole deliberation in your hands and you persisted the conclusion.

The deliberation you threw away

Consider what a rejected-set actually carries, and you will see why its absence is not a minor gap. A decision's defensibility lives largely in the comparison. Not in the chosen outcome standing alone — any outcome can be asserted — but in the relation between the outcome and the ones it beat. Was the winner clearly ahead, or did it edge out the runner-up by a hair? Was there a reasonable alternative that scored nearly as well and would have served the affected party far better? Was there an obvious option that the system never evaluated at all, so that the choice was made inside a field that had already been narrowed by something nobody recorded? None of these questions can be answered from the winner alone. All of them are the questions that decide whether the decision was sound.

So the record that keeps only the output cannot distinguish between two decisions that could not be more different in kind. One was close — two strong options, a real weighing, the chosen one ahead by a margin so thin that a slightly different input would have flipped it. The other was a walkover — one option dominant, the rest not seriously in contention. From the output, these look identical: same verdict, same clean line in the log. But they are not the same decision, and an affected party has every right to know which one happened to them. A close call presented as though it were obvious is a small deception, and a record that cannot tell the two apart is complicit in it.

And there is one question the winner-only record can never answer, which is the most natural question anyone on the receiving end will ask: why not the other outcome? Why not approved. Why not the route that would have cost me less, why not the action I would have preferred. The system had an answer to that question at the moment it decided — it had the other option in hand and a basis for setting it below the one it took. But if you did not write the comparison down, that answer is gone, and all you can offer in its place is the outcome repeated more firmly. Which is not an answer. It is the refusal of one.

You cannot tell a hard-won decision from a foregone conclusion by its outcome; only the alternatives it beat can tell them apart.

Record the option set and the ranking

Concretely, then, capture three things at the point of decision. First, the option set the system actually evaluated — the alternatives that were genuinely in contention, not just the one that prevailed. Second, the basis on which each was ranked or rejected: the score it received, or the rule that disqualified it, or the reason it fell short — whatever the system in fact used to order them. Third, how close the decision was: the margin between the winner and the strongest rejected option, so that a reader can see whether this was a clear call or a near-tie. Persist these bound together with the outcome, as one account of one choice, written before the outcome is returned rather than reconstructed later from a system whose state has moved on.

What this buys you is a decision that can be examined as a choice rather than merely asserted as a result. The affected party, the reviewer, the auditor a year later — each of them can now see not just what the system did but the field it did it within, and can test the choice against its own alternatives. Was the runner-up actually better on the terms that should have mattered? Now the question has an answer that lives in the record instead of in someone's memory or the system's current, drifted configuration. This is the counterfactual a contested decision needs — the thing that lets a rebuttal say not merely I disagree but here is the option you should have taken, here is where it ranked, and here is why your own basis for rejecting it does not hold. A decision you can argue with on its own terms is a decision that can be defended on them, too.

The honest limits

This directive has edges, and the discipline is in respecting them rather than overreaching past them. You cannot record infinite alternatives, and you should not try. The instruction is not to fabricate an exhaustive list of every conceivable option and pretend the system weighed them all — that is its own kind of falsification, a deliberation staged after the fact for the record. Capture the option set the system actually evaluated, honestly bounded. And where it matters, note what the system did not consider — because a reasonable alternative that never entered the field is often the most important thing about a decision, and a record that silently omits it presents a narrowed choice as though it were the whole of the possible. The absence, acknowledged, is more truthful than a fuller-looking list that hides it.

The other edge is stakes. Not every decision earns a full rejected-set, and pretending otherwise buries the consequential decisions under the volume of the trivial. A system making thousands of low-stakes calls a second does not owe each one a recorded deliberation; the depth of alternative-recording should scale to what is at risk in the decision. For the routing choice no one will ever contest, the winner may be enough. For the decision that changes a person's access, their standing, their record, the comparison is exactly what you will be asked to produce — and exactly what you will not be able to reconstruct if you did not keep it. Spend the recording where the stakes are, and spend it fully. The point was never to log more. It was to make the decisions that matter examinable as the choices they actually were.

— Dispatches · Summit Cognitive