RLCD (Reinforcement Learning for Calibrated Decisions)

Reinforcement Learning for Calibrated Decisions (RLCD) is a training method TypeSafe AI says it used to build Jev, its first "System One" model. The company announced Jev in a September 2026 blog post by cofounder and CEO Diogo Almeida. Reinforcement learning from human feedback (RLHF), a method Almeida co-developed at OpenAI for InstructGPT, trains a model to produce outputs that human raters prefer. RLCD instead optimizes a model's output probabilities so that they are calibrated: among all the predictions the model assigns a given confidence level, such as 90%, that fraction of them should actually turn out correct.

Calibration in this sense says nothing about whether any single prediction is right. It only means the model's stated confidence tracks its accuracy in aggregate across many predictions. TypeSafe treats this as distinct from a model simply being correct or incorrect. Because Jev's outputs are constrained to a fixed schema, a malformed or out-of-schema answer counts as a separate failure mode from a well-formed but wrong one. RLCD targets the confidence attached to well-formed answers specifically, aiming to make it trustworthy enough that software can act on it automatically above a chosen threshold.

As of its announcement, RLCD has been described only in TypeSafe's own blog post and subsequent press coverage. No peer-reviewed paper or technical report documents its exact mechanism, so the loss function, reward model, and calibration procedure remain undisclosed. Commentators have noted this distinction, describing the architecture, sampler, and training method as company claims rather than verified technical results pending a fuller technical writeup.