This Japanese tutorial explains TypeSafe's Jev decision model, distinguishing its selection and scoring paradigm from text-generating LLMs. It details stateless criteria evaluation, sensory parsing pipelines, confidence metrics, and practical interactive logic design.
Original by ざすこ (道草_雑草子)Getting startedBeginner12 min 36 secPublished
Jev acts as a specialized decision model that selects options, scores rubrics, or yields probabilities from predefined candidate choices rather than generating freeform conversational prose.
Because Jev is text-based and completely stateless, it retains no prior session memory, requiring developers to provide explicit state, questions, criteria, and pre-extracted modality data on each call.
Form-fitting structured output does not guarantee factual correctness; human-in-the-loop workflows must route low-confidence scores or high-stakes edge cases back to human review.
Worth knowing
Jev only evaluates text representations and does not execute code, inspect raw multi-modal files directly, or guarantee correct decisions simply because outputs conform to target schemas.