The decision that never ends

Accountability was built for decisions that happen once. Someone approved the loan on a Tuesday. Someone set the price. Someone assessed the risk and wrote it down. Whatever went wrong afterward, there was a discrete act to return to — a moment you could name, reconstruct, and argue about. The whole apparatus of holding a decision to account assumes this shape: that a decision is a thing that occurred, at a time, and that pointing to it is at least possible. Much of what automated systems now do refuses to hold that shape at all.

Consider a system that does not decide once but re-decides continuously. It adjusts a feed as new signals arrive. It nudges a price up and down through the day. It recomputes a risk score every time a fact changes, a hundred times an hour, silently, without any of the recomputations announcing itself as a decision. Ask when such a system decided anything about you and the question dissolves in your hands. It decided constantly. It is deciding now. There is no Tuesday to return to, because every instant is a fresh instant of deciding and none of them is the one.

This is not a marginal case. It is increasingly the ordinary one. The continuous decision — the price that breathes, the score that drifts, the ranking that re-sorts under you as you read it — is the native form of automated judgment, and it is the hardest possible form to hold to account. Not because it is complicated, though it may be, but because accountability has no handle on it. You cannot cross-examine a process. You can only cross-examine an act, and the process has been carefully arranged, though usually not on purpose, to never quite commit to being one.

Why a stream cannot be held to account

The difficulty is structural. To hold a decision to account, you have to be able to say what was decided, when, and on what basis — and each of those requires an edge, a boundary that separates the decision from the ceaseless activity around it. A discrete decision comes with its edges built in: it started when the file was opened and ended when the form was signed. A continuous decision has no edges. It is all middle. There is no opening and no signing, only an endless adjustment that never resolves into a moment you can hold up and examine.

And so the affected person is left in a peculiar position. Something was done to them — the price they paid, the rank they received, the score that shut a door — but there is no decision to point at, only a system that was, at the relevant instant, doing what it always does. The harm is specific. The act is diffuse. When they ask what happened, the honest answer from inside the system is that nothing in particular happened; the process simply continued, and their outcome fell out of it the way a temperature falls out of the weather. That answer is true and completely useless, which is the signature of a decision that has escaped accountability by never quite being made.

You cannot hold a river to account for where it deposited the silt. You can only hold it to account if you agreed, in advance, which bends would be measured.

Imposing an edge on the stream

The answer is not to pretend the stream is discrete when it is not. It is to impose an edge on it deliberately — to decide, in advance, which moments in the continuous process count as decisions worth recording, and to mark them as they pass. The record's job, for a continuous system, is not to log everything; logging everything is its own way of recording nothing, because a trace with no structure is as unnavigable as no trace at all. The job is to draw the boundary the stream will not draw for itself: to say that when the price crossed this threshold, when the score fell below that line, when the ranking demoted this item past the point where it would be seen — those were decisions, and here is what each one looked at.

This is a design act, and it has to happen before the fact. You cannot go back to a stream after something has gone wrong and manufacture the edges retroactively, because the whole point of an edge is that it was drawn at the moment that mattered, capturing what the system saw then rather than what it can reconstruct now. A record that marks the consequential moments as they occur turns an unaccountable process into a sequence of accountable acts. It converts the river into a series of readings taken at agreed bends — not the whole flow, but the moments that carried consequence, fixed as they went by.

The discipline of admissible reality insists on exactly this: that a system be able to produce, on demand, an account of the specific decision that affected a specific person, even when the system itself experiences no discrete decisions at all. The account is not a fiction imposed on a formless process. It is the deliberate marking of the moments the process would otherwise let slip — the threshold crossed, the line breached, the outcome fixed — so that a continuous decision can still answer to a discrete question. The person harmed does not need the whole stream. They need the one bend where the silt was deposited on them, measured because someone decided in advance that it would be.

Absent that, the continuous decision remains the perfect vehicle for unaccountable power: consequential in its effects, formless in its structure, always in progress and never quite made. The record is what forces it to stop, at least at the moments that count, and say what it just did.

— Dispatches · Summit Cognitive