Overview
- Octomind rolled Jev routing into version 0.54.0 and enabled it by default, using the model to replace many small agent judgments such as skill routing, output scoring, verifier checks, memory writes, and planning decisions.
- Jev, from TypeSafe AI, exposes three typed primitives—Choice (probabilities per option), Score (a probability distribution across ordered levels), and Noul (a yes/no probability)—and returns calibrated probability distributions instead of generated text.
- Adopters report large per-decision savings: Octomind measured per-answer latencies of about 0.5–1.3 seconds and per-call costs roughly $0.00001–$0.0001, while TypeSafe advertises tens-to-hundreds of milliseconds and bills input at $0.042 per million tokens.
- Early implementers documented four practical pitfalls to avoid: use argmax of returned probabilities for Score buckets rather than rounding the numeric score, treat the provided confidence as distribution peakedness not absolute truth, avoid splitting synonyms across options, and remember omitted options cannot be chosen.
- Engineers recommend using Jev as a cheap veto or filter—escalating uncertain cases to full LLMs—and validating calibration on your own data because TypeSafe’s calibration is measured on its datasets and may not hold in every domain.