Our summary
Francesco Cerizzi wrote this Medium post, published on 4 November 2026. It explains what a decision model is and how open ones on Hugging Face read a state plus typed questions and return one probability per answer, then compares seven designs with Jev 1.14 and OpenAI's Decisions API.
On his 200 routine decisions the four larger open models were 93% to 98% right and answered in 106 to 194 ms on one GPU, similar to Jev 1.14 and the Decisions API. On long records the spread was wide, from 53.6% to 97.0%, with the two models that read the record in both directions leading.
It is one author's test on his own 496 decisions, run on rented GPUs and AWS Lambda, so latency includes network time for the hosted services. Calibration did not always carry over: Jev's calibration error rose on long records. The author points to the Jev Decision Index for choosing a model.
Key takeaways
- Use a decision model wherever your code needs one closed answer with a probability, such as routing, guardrails or triage.
- Measure calibration on data that looks like yours, because a temperature fitted on one kind of record may not carry over to another.
- Check long, rule-heavy records separately, because that is where the designs differed most.
