A standard for decisions a machine made.

The problem

Machine-made decisions have outrun our ability to account for them.

In a modern organization, the consequential decisions are no longer made by one person in a room. A human asks a system a question. A model returns an answer. Someone trusts it, or edits it, and acts. The decision lands on a real person — an applicant, a patient, a claimant, a defendant — and then the chain that produced it disperses.

Try to reconstruct it afterward. What inputs did the system actually see. Which version of the model answered. What did it return before a human softened it. Who held the authority to override. In most organizations the honest answer to each of those questions is the same: we did not keep it, or we kept some of it, in different places, in formats that do not line up.

We have built decision pipelines that produce consequences but do not produce evidence. For a while that did not matter, because no one ran the review, so no one discovered they could not. That time is ending. The stakes are no longer low, and the questions are getting louder.

We kept the permission to act. We quietly threw away the proof.

The shift

From is it true to would it be admissible, and on what record.

For most of history we judged a claim by asking whether it was true, and we answered that by reading the speaker: who said it, how sure they sounded, how costly it would be to check. Those proxies held because looking authoritative was about as expensive as being authoritative. A machine severs that bond. It produces fluent, confident, internally consistent output with no relationship between how sure it sounds and how sound it is, attached to no maker who can be summoned and made to answer.

So the question has to move — off the claim and onto the record around it. Not is this output correct, but is this a thing I am entitled to rely on, given how it was produced and what it can show about its own production. The word for that is admissible, and it is borrowed deliberately. Courts long ago gave up the fantasy of directly verifying the truth of everything put before them; they cannot rerun the crime. So they built, over centuries, a discipline for deciding which claims are even allowed to count — where the evidence came from, who offers it, whether the other side can test it.

Admissibility is not a lower standard than truth. It is the standard you adopt once you admit you cannot establish truth by inspection, and must instead govern the conditions under which a claim earns the right to be relied upon. A wrong decision can still be admissible — and that is the point, because an admissible wrong decision can be caught, while an unaccountable right one cannot be trusted and an unaccountable wrong one cannot be found.

The four properties

A claim is admissible when it arrives with enough of the right record to be relied upon and, above all, to be contested. That record rests on four properties. A record missing any one of them is incomplete.

What we are building

One argument, carried in four forms.

This is not a doom statement; it is a design statement. The record is not something you bolt on at the end. It is built into the moment the decision is made, or it does not exist at all. The work below is the body of work pursuing that standard — and the standard can be taken, built, or rejected by anyone, which is the only honest way to offer it.

An invitation

We are setting the standard for what machine-made authority will be allowed to get away with — mostly by default.

To the builders: you will be the ones answering when a decision is challenged, and it is cheaper to be early than to retrofit under pressure. To the leaders: when a regulator, a court, or an affected person asks why a decision was made about someone, the organization that deployed the system answers — not the model vendor three layers up. And to the affected: you deserve a record you can use to push back, not a polished account you can only resent.

The ask is narrower than buy this. It is: stop treating the record as optional where the stakes are real. Make the standard the default before its absence becomes the default — because one of the two is going to set, and only one of them leaves us able to ask the question that started all of this and still get an answer worth having.

— Dispatches · Summit Cognitive