What is a decision model? A close look at Jev, Strands Decider, Clef and more

lmoroney · x · 2026-10-03

Laurence Moroney published a systematic guide to decision models, the new category kicked off by Jev on Sep 15 and now including Tev1, Nimble, Strands Decider and Clef.

What they do: a decision model takes a state (support ticket, JSON blob, proposed tool call) plus your questions and candidate answers, scores every answer and returns typed results with probabilities. Since it can only pick from your list, output always fits your schema. The dominant API defines three question types: Choice (pick one, Jev supports up to 255 options), Score (distribution across an ordered scale), and Noul (probability a condition is true).

Use cases: routing requests to the right model/queue, policy checks, deciding whether an agent calls a tool, ticket triage, cheaply judging another model's output.

Two designs: keep a pretrained LM body and swap the text head for a small pointer head over options — AWS Strands Decider's head is just over 1M parameters — or the approach taken by Cloudflare's Clef. The guide also covers reading confidence numbers and picking a model for your project.

Original post →

More from coding & agent

coding & agent channel →