What Is RLCD? The Secret Behind Jev

RLCD is a formalization of reward modeling that reframes scalar scores as explicit, calibrated multiway preferences. It traces the evolution from scalar reward functions to pairwise preference models (Bradley-Terry/PPRM) and then to the Plackett-Luce multiway choice model, which normalizes utilities across K alternatives into a softmax distribution. RLCD supplements this choice likelihood with probability calibration so reported probabilities carry empirical meaning: high probability should correspond to high observed accuracy. Calibration is enforced and measured with proper scoring rules (notably the Brier score) and can be tuned post-hoc via temperature scaling, separating ranking quality (which candidate is best) from confidence quality (how often that choice is right).

Jev operationalizes RLCD by making the evaluator the runtime product rather than an internal reward component. Instead of generating candidates and scoring them post hoc, Jev accepts state and candidate schemas and returns calibrated decision distributions and confidences through three typed primitives - Noul (binary truth), Choice (Luce distribution over alternatives), and Score (ordered-level distribution). This design enables parallel inference, clearer decision policies (execute, defer, escalate based on calibrated confidence), and concrete evaluation via reliability, resolution, and uncertainty metrics, producing testable predictions about when automated actions are safe to take.