Jev introduces a new shape of LLM - System One, aka Decision Models

What it claims

Jev, from TypeSafe AI, takes text or semi-structured input and returns only floating point numbers: probabilities, scores and confidence values. TypeSafe frames it as "unstructured state in, typed probabilistic decisions out." Input costs $0.042 per million tokens and output is free. Willison notes that this undercuts GPT-5 Nano at $0.05 per million. He finds it useful for anything that reduces to classification, including spam detection, label suggestion, prioritization, ranking and search reranking.

How to use it

Request shape

• You send one "state" plus one or more questions.
• The state can be a string, an array of strings, or a set of name-value pairs, such as an article or a customer record.
• You can include as many questions as fit in the context window. They are evaluated in parallel, so many questions take about as long as one. To cut latency, batch every question about a record into one call.

Return values by question type

• Noul (yes/no): you pose a statement and get a 0–1 confidence that it is true.
• Choice: you get a confidence score plus a probability distribution across all provided options, not just one pick.
• Score: you supply a sequence of numeric levels, each with a description, and get a float somewhere along that range.

Reranking pattern Willison uses

1. Fetch about 100 candidates cheaply, for example with BM25.
2. Have Jev score each one for relevance to the query.
3. Re-sort by that score.

From the LLM CLI Install the llm-typesafe plugin (released 22 September 2026). A basic call:

llm -m jev 'Please refund my last payment.' \
  -s 'Does this message explicitly request a refund?'

The plugin README covers the other query types.

Open-weight alternatives

• Kev recreates the idea on Qwen 3.5 at 0.8B, 4B and 9B sizes.
• JevBench has appeared for comparing "Jev-class decision models." The post gives no scores from it.

Limits

• Known weak spots: TypeSafe's own Jev 1.13 "jaggedness" documentation says the model is currently not great with numbers, dates or adversarial content. Check this before using it on anything with those inputs, or where users can manipulate the input.
• No explanation: you get a number and nothing else. Even an unreliable rationale, which a text LLM can give, is unavailable, so you cannot see which signals drove a spam call.
• Bias risk: Willison specifically warns against using Jev to rank job applicants. In one informal test he asked "Good city?" about every Bay Area city; Jev rated Cupertino top and East Palo Alto bottom.
• What to do about it: his recommendation is heavier reliance on evals and structured experiments than with ordinary LLMs. At these prices, thousands of probe prompts cost a few cents.
• Maturity: the post was written less than a week after launch. Its "community uses" are novelties (a chat model built one symbol at a time, left-pad, 2048), not evidence of production quality. The post contains no benchmark results.