TypeSafe AI Ships Jev. It Returns Decisions, Not Paragraphs. The Typing Is The Point.
TypeSafe AI released Jev, a system one model that returns typed, calibrated decisions with probabilities instead of text output. It costs $0.042 per 1M input tokens. The researcher behind the project deserves credit, though the source does not name them.
This demonstrates the principle of output constraint as a feature, not a limitation. The mechanism is typed decision returns: instead of generating free text that you then parse, the model hands you a structured decision with a calibrated confidence score. You should understand that most production AI failures come from interpreting unstructured output. Jev eliminates that failure mode by construction.
TypeSafe AI built and released Jev. The source credits the project's researcher but does not name them.
- Open ChatGPT or Claude and ask it to decide whether you should bring an umbrella today, but instruct it to respond ONLY in valid JSON with fields for decision (yes/no) and confidence (0-1).
- Try the same prompt without the JSON constraint and notice how much harder the output is to use programmatically.
- Ask the model to assign a numerical confidence to its decision and compare that to its actual accuracy. You will find calibration is harder than it looks.