The record and the ritual

Consider the older way of making a consequential decision. A person sat at a desk, read the file, and then did something slow and slightly theatrical: they signed. Perhaps a witness watched them sign. Perhaps an oath had been sworn earlier, or a form countersigned, or a seal pressed into wax that no longer served any structural purpose except to say that this moment was not like the others. We have grown used to reading these rituals as friction — as the bureaucratic residue of institutions that had not yet learned to move quickly. But the friction was doing work, and the work it did was more important than the ritual ever let on.

The ritual slowed the decider down. It inserted a pause between intending a thing and having done it, and in that pause a decision could be reconsidered, or at least recognized as a decision. It marked the moment as consequential, which is to say it told the person that what they were about to do would be remembered, and that they might one day be asked about it. And, almost incidentally, it produced a record. The signature was a fact that outlived the signing. The witness could later be asked what they saw. The form, once filed, became a thing that existed in the world independent of anyone's memory of the afternoon it was made.

None of that was the point of the ritual, exactly. The point was solemnity, or authority, or the reassurance of doing things the way they had always been done. The record was a byproduct. But it was a reliable byproduct, and a whole architecture of accountability came to rest on it without anyone quite deciding that it should. We could ask who signed because signing left a mark. We could reconstruct a decision because the ceremony of making it had scattered evidence in its wake.

What automation removes without announcing it

An automated decision has no ceremony. There is no desk, no pause, no witness, no seal. The decision happens at a speed and a scale that would make ritual absurd; you cannot press wax into a million approvals a second. And so the record that ritual produced by habit simply does not form. Nothing scatters evidence in the wake of an automated decision unless someone has arranged, in advance and on purpose, for it to do so. The default state of an automated decision is not a light record or a partial one. It is no record at all.

This is easy to miss, because the absence looks like efficiency. We tell ourselves we have removed the friction and kept the function — that the decision still gets made, only faster and cleaner. But we removed more than friction. We removed the pause in which a decision could be recognized as one. We removed the mark that outlived the making. We removed the witness who could later be asked. The function that survived is the narrow one, the arrival at an outcome. The functions that quietly died are the ones that made the outcome answerable afterward.

Ritual never advertised itself as record-keeping. That is exactly why we did not notice when its record-keeping went away.

There is a temptation to treat this as nostalgia, as though the argument were that we should reintroduce seals and oaths for their own sake. It is not. The ceremonies were arbitrary; the wax served nothing. What matters is not the form of the ritual but the byproduct it happened to yield. The mistake is to assume that because the ritual was inessential, the record it produced was inessential too. The ritual could go. The record could not — and it went anyway, because it had never been anything more than a side effect of a ceremony we were right to abandon.

Rebuilding on purpose what once happened by habit

The task, then, is not to slow decisions back down. It is to reproduce deliberately what the pause and the signature and the witness produced by accident. An automated system that wants to be accountable has to build, as an explicit feature, the thing that ritual used to throw off for free: a mark of the decision that outlives the deciding, made at the instant the decision is made, carrying enough of what was in front of the system to let someone later ask what happened and get an answer that is not merely the system's own account of itself.

This is what a real decision record is for. Not to add friction back — a record can be produced at machine speed, and a good one is — but to restore the byproduct after the ceremony that used to yield it is gone. The record does what the witness did: it lets a later party ask what was seen. It does what the signature did: it fixes a moment that would otherwise dissolve into the general blur of the system's operation. It does what the pause did, not by slowing anything down, but by making the moment recognizable as a decision after the fact, so that it can be treated as one when someone comes asking.

The doctrine that ties this together — call it admissible reality, the discipline of building systems whose accounts of themselves can actually be relied upon — begins from this recognition. What ceremony once gave us by accident, we now have to engineer on purpose, or do without. A system that removes the ritual and does not replace the record has not become more efficient. It has become more powerful and less answerable at the same time, which is the one combination we have the least excuse for choosing by default.

So the question to ask of any automated decision is not whether it moved fast enough. It is whether, in moving fast, it left behind what the slow way left behind without trying. If it did not, the efficiency was purchased with something we did not mean to spend — and the bill, as always, arrives later, when someone asks what happened and finds that nobody made a record because nobody performed the ceremony that used to make one for them.

— Dispatches · Summit Cognitive