A typed decision model returns a choice and a confidence, not text. Govern it like any deciding system. The standards name six places its authority sits.
A typed decision model takes a record of the case and a set of questions with fixed answers, and returns a choice, a score, or a probability with a confidence margin. It writes no text and gives no reasons. The first widely adopted product in the class, TypeSafe AI's Jev, was released in September 2026 and marketed as hundreds of times faster and cheaper than a language model for routing, triage, tool gating, and approval checks.
The method needs nothing new for it. The unit of governance is the consequential decision, whatever component made it. What changes is the price, and so the number of decisions. When a judgment costs a fraction of a cent and a few hundred milliseconds, institutions put one at every branch, and authority moves into places that do not look like decisions.
The standards name six of those places: the threshold that decides between acting and asking, the content in the state the model reads, the queue a reviewer is left with, the hops that shape a decision without making it, the reasons nobody produced, and the rate at which decisions outrun the capacity to answer challenges.
A hypothetical case. A support platform puts a decision model in front of its refund agent. Refunds scoring above 0.8 run unattended and the rest go to a reviewer. The 0.8 lives in a configuration file, was tuned on the previous model version, and the reviewer only ever sees borderline cases. The standards ask for the 0.8 to be recorded as a policy, tied to the model version it was tuned on, so a new model version reopens it. They also ask that a declared share of the refunds that ran unattended go to review, so a confident error has somewhere to be seen.
Mechanism, drawn
The model returns a choice and a confidence, not text. Its authority sits around the call, in six places the standards name.
In practice
The number that sends a case to automatic execution or to a person is where authority is conferred. Record it as a policy that names the labeled cases, model version, question schema, and option order it was set against. If the model version is not pinned, the threshold may already be running on a model it was not tuned for.
A gate that reads emails, pages, and tool outputs can be argued with by them. A forged note claiming the user already approved has moved a real gate across its threshold. Authority comes only from the records that grant it, never from the material the gate reads.
A queue fed only with low-confidence cases cannot catch a confident error. Treat the routing threshold as part of how review is set up, so changing it reopens that setup. Send a declared share of the decisions that skipped review back into it.
A classifier that ranks, filters, or routes what the deciding model sees shapes every decision it touches. A router that picks the model delegates every case it routes. Both belong on the list of components behind each decision. A router is listed with every model it could pick, as if each ran on every decision.
A model that gives no reasons invites a language model to write one afterward. That text is an account of a decision it did not make. Label it as an account, and have the decision record say what the decision rested on: the policy, the threshold, and the inputs.
At machine speed, one class of decisions can produce more errors in a day than the people answering challenges can handle in a year. Decisions issued, times the error rate, may not exceed the number of challenges those people can answer in time. A class past that cap is treated as uncontestable in practice.
Where the standards use it:
A typed model cannot return an answer outside its schema. It can still return a valid, confident, wrong one. Schema validity is a property of the output format, and the method governs the decision the output becomes.
Not on its own. The vendor's own documentation describes confidence as a margin, not a probability that the answer is right, and calibration measured on someone else's workflows is not evidence about this deployment. Law III couples authority to evidence the institution holds. Confidence does not authorize.
Because the reasons would describe a plausible decision, not the one that was made. A challenge depends on the reason being something the person it decides about can dispute and that would change the outcome if wrong. An account written by another component meets neither test.
No. The method takes no position on which component decides. Typed output can make governance cheaper, because the choice, the distribution, and the confidence are exactly the fields a decision record needs. The risk is in treating a decision as plumbing because it was cheap.