The minutes, not the verdict

A town council meets on a Tuesday evening to decide where a stop sign should go. There are perhaps nine people in the room and one item on the agenda. And yet, before the meeting begins, someone is appointed to keep the minutes. Who moved the motion, who seconded it, what objections were raised, which alternatives were considered and set aside, how each member voted, and on what authority the council was acting at all. We do this for a stop sign. We have done it, in some form, for every serious deliberative body for centuries, because we learned long ago that a decision is not the same thing as an account of a decision — and that the second is often the more important of the two.

Consider what we actually demand of the institutions we take seriously. A court does not simply announce a verdict; it produces a record — the evidence admitted, the procedure followed, the reasoning entered into the opinion, and, crucially, the dissents. A board does not merely resolve; it minutes the resolution, the abstentions, the disclosed conflicts. A legislature keeps a journal. A standards committee publishes the comments it received and how it dispositioned each one. In every case the verdict is the smaller artifact. The larger one is the recorded path that produced it, and it exists for a specific reason: so that someone who was not in the room, and who may not trust the people who were, can judge whether the decision was reached legitimately.

That is the function the minutes perform. They are not a courtesy and they are not a transcript for nostalgia's sake. They are the instrument by which an outsider can audit a process they did not witness and could not have witnessed. The verdict tells you what was decided. Only the minutes tell you whether it was decided in a way that deserves your acceptance.

What the minutes are for

It is worth being precise about why the record so often matters more than the outcome. A verdict, taken alone, is unfalsifiable. "Approved." "Denied." "Liable." There is nothing in the word itself you can grab hold of and test. You can resent it, you can appeal it, but you cannot inspect it, because there is nothing inside it to inspect. The minutes are what convert an outcome from an assertion into something with a structure — a thing made of steps, each of which could in principle have gone another way, each of which can be examined for whether it followed the rules it was bound by.

This is also why we keep the dissents. A unanimous-looking verdict with the disagreement scrubbed out is not a cleaner record; it is a falsified one. The points where a decision could legitimately have gone the other way are exactly the points an outsider most needs to see, because they are where legitimacy is actually tested. A record that hides its own contested junctures has hidden the only places worth auditing.

A verdict tells you what was decided. The minutes are the only thing that tells you whether it was decided in a way that deserves your acceptance.

So when we built machines to make decisions at scale, the question that should have been asked first was not how accurate they were. It was: where are the minutes? And the answer, almost universally, was that there weren't any. The systems were allowed to do the one thing we have never permitted a serious institution to do — keep the verdict and throw the deliberation away.

What minutes mean for a machine

The objection is obvious: a credit model does not deliberate the way a council deliberates, so the analogy must be loose. But the analogy is tighter than it first appears, because "minutes" was never really about the meeting. It was about reconstruction. The minutes let you rebuild the path from inputs to outcome closely enough to judge it. A machine decision has exactly such a path, and we can name its parts.

The evidence actually consulted — not a description of the kind of data the system generally uses, but the specific inputs that were in front of this decision at the moment it was made. The rules active at the time — the policy, the thresholds, the model and its version, frozen as they stood at that instant rather than as they have since been amended. The steps taken — the order in which things were evaluated, because order frames meaning here as much as it does in a courtroom. And the junctures: the points at which the decision was close, where a slightly different input or a slightly different threshold would have produced the opposite outcome. Those are the machine's dissents, and they are the most important thing the record can carry.

A system that emits only "denied" has given you a verdict. A system that can hand you those four things has given you minutes — a record from which the deliberation can be reconstructed and, if necessary, replayed against the inputs to confirm that the stated path actually produces the stated result. That is not a soothing explanation generated after the fact. It is the deliberation itself, preserved.

Admissibility is the demand for minutes

This is what admissibility actually asks for, stripped of jargon. It is not a demand that machines explain themselves in plain English, which they can do persuasively and falsely. It is the older, narrower demand we have always made of serious institutions: keep a reconstructable record of how you decided, not merely a stamp of what you decided. Keep the minutes. A Decision Receipt is, in this light, not an exotic new artifact at all. It is the attempt to make minutes the default output of an automated decision rather than a favor extracted afterward, through a lawsuit, from a system that was never built to surrender them.

And that is the irony worth sitting with. We require minutes of the council deciding where to put a stop sign. We require them of the zoning board, the co-op committee, the disciplinary panel of a small professional society. We have decided, as a civilization, that any body whose decisions touch other people owes a recorded account of how it reached them. Then a system denies someone a loan, or flags them for review, or declines their claim — decisions with rather more weight than a stop sign — and we have, for a while, accepted silence. An outcome, and at best a log confirming that an outcome occurred.

There is no principled reason for the exemption. The machine decision is not less consequential than the council's; it is more. It is not harder to record; in some ways it is far easier, since the path is already structured data the moment it is made. The exemption survived only because nobody insisted, and because the systems were shipped before the question was asked. The correction is not to make machines more eloquent. It is to hold them to the standard we already hold a roomful of volunteers deciding about a stop sign — to require that they, too, keep the minutes, and surrender them on demand to the person the decision was about.

— Dispatches · Summit Cognitive