Kev: Tiny Jev-like family of decision models built on top of Qwen3.5

Kev is a small family of Jev-style decision models built on Qwen3.5 that you can run or train locally. It exposes a System One-compatible API (noul, choice, score question types) so TypeSafe clients can query a single input with multiple isolated questions and receive calibrated probabilities, confidences, and scores rather than a single label. Models (0.8B, 4B, 9B) ship with training code, evaluation suites, and a web playground for experimenting with question isolation and option order. Serving is designed for CUDA and Apple Silicon (bf16), with quickstart instructions using uv, example curl/JSON requests, a Python SDK snippet, and an interactive Node-based playground including a chess demo. Responses include per-question probabilities, confidence, token usage and latency.

Evaluation tables report accuracy and Brier scores on held-out training sources and on new-source transfer sets. Kev-4B and Kev-9B show strong accuracy (Kev-4B ≈0.877 dev / 0.870 test; Kev-9B ≈0.876 / 0.873) and improved calibration versus smaller generations; Kev-9B trails a hosted Jev baseline by about 4.5 points on new-source development data. Controlled generation comparisons (Qwen3 → Qwen3.5 bases) show notable gains: Kev-9B is ~7.3 points ahead of its Qwen3 predecessor on the test set with a 0.08 lower Brier, and smaller models also improve. All weights and model cards are published with SHA-256 checksums under an Apache-2.0 license.