What was covered
- The rule behind the talk. Code calculates, JEV judges, the LLM reasons, and anything uncertain goes to a human.
- The problem. Most decisions do not need text. With an LLM you ask it to pick an option, parse and validate the JSON, retry when it breaks, pay for output tokens and wait seconds. A decision model takes a state and a typed question, with the options defined in advance, and returns a probability per option in a single pass.
- JEV. Made by TypeSafe AI, launched on 15 September 2026 as a closed API at $0.042 per million input tokens. Choice picks one option out of up to 255 and returns a probability for each. Score returns the probability that a statement is true, where 0.5 means “can’t tell”.
- Against an LLM. Typed values instead of text to parse, output that is free, latency of 70 to 500 ms as reported by TypeSafe, and nearly deterministic results. It cannot reason or write, and it is weak at arithmetic and dates, which belong in code.
- Architecture. Code owns the flow. JEV, an LLM or an open model sits behind the same contract. The policy is a pure, testable function, and anything below a confidence threshold goes to human review.
- Decision Lab. A FastAPI app, with a static UI and a Docker image, that runs JEV against up to four LLMs on your own tickets and reports accuracy, paging, median latency and cost per contender. Keys are typed into the page and the server stores none.
- Our own run, 40 tickets. JEV: 40 of 40 correct, 306 ms median, $0.020 per 1,000 tickets. gpt-6-luna via LiteLLM: 39 of 40, 851 ms, $0.039. qwen3.8-27b via LiteLLM: 39 of 40, 10.1 s, no price set.
- What the run supports, and what it does not. JEV answered 2.8 times faster than gpt-6-luna at about half the cost. But 40 of 40 against 39 of 40 is a single ticket, and forty tickets are a demo, not a benchmark.
The wider picture
- A valid answer can still be wrong. Check Point showed that adding text that looks like evidence to a report flipped a fraud assessment to low risk in all nine configurations tested, at around $0.50 per successful attack. Telling the model to ignore the document did not help. Keep a person on anything irreversible.
- A category, not a product. In 24 days the market went from one vendor to at least eight. Several copy JEV’s request format, and prices fell to as low as $0.02 per million. The gap is now quality, modality and where it runs.
- When to use which. Code for arithmetic, dates, permissions and limits. A decision model for classifying, routing, scoring and filtering. An LLM to reason, plan and write. A self-hosted open model when data cannot leave your network. A human for decisions about people.
What came up
- Trying it on an existing project. The room discussed testing JEV or an open alternative against an ageing routing algorithm. Bruno will build a small proof of concept to make the case.
- Small models. Felipe proposed a website for finding interchangeable language models of about three billion parameters.
Takeaway
Keep the decision layer swappable, and measure on your own tickets: label a few hundred cases, run Decision Lab on your own data, then set thresholds and test in shadow mode. These figures were four weeks old when presented.