Abstract

Autonomous agent loops spend substantial compute evaluating discrete, bounded choices: tool dispatch, intent routing, and policy safety checks. Kai decouples bounded judgment from open-ended text synthesis, emitting temperature-calibrated probability distributions across typed candidate sets (choice, noul, and ordinal score) via a factorized cross-attention reader. On the frozen 12-suite decision harness (2c74e2d), kai-1 achieves 0.857 mean accuracy vs 0.780 for Jev 1.13 (leading on 11 of 12 suites) and 0.093 mean ECE vs 0.139, while composing with Nio for deterministic provenance and policy verification.