Training
RLCD
2026ActiveUpdated: 17 September 2026Published
Key
innovation
Training that optimizes not text but calibrated, "epistemically honest" decision probabilities — probabilities tuned to real outcomes rather than to human preferences.
Category
Training
Abstraction level
Pattern
Operation level
Post-trainingTraining
Use cases
Training decision models (System One / Jev)Automating in-software decisions with confidence thresholdsRouting cases to a human or a reasoning model based on calibration
How it works
Instead of a reward signal based on human preferences (RLHF) or verifiable task rewards (RLVR), RLCD uses a signal that rewards calibration: the probabilities assigned to decisions are optimized against real outcomes so that, over large groups of predictions, they match the observed frequency of correctness. The result is models that return typed decisions (e.g. a choice, a score, a boolean) together with a trustworthy confidence level.
Problem solved
Language models trained for human preferences tend to be overconfident and poorly calibrated, which hinders automated, safe decisions in software; RLCD aims to yield probabilities that reflect genuine uncertainty.
Evolution
Original paper · 2026 · TypeSafe AI (blog) · Diogo Almeida
Introducing System One Models and Jev
Diogo Almeida
2026
RLCD and System One models (TypeSafe AI)
Inflection pointRLCD announced alongside the Jev model in September 2026.