The name sits alongside two older acronyms. RLHF, reinforcement learning from human feedback, trains a model toward responses human raters prefer. RLVR, reinforcement learning from verifiable rewards, trains it toward outputs a program can check as correct. RLCD swaps the target again: the reward is tied to whether the stated probability matches how often the answer turns out to be right.
That target is what calibration means. A model trained this way should be right on roughly 80% of the answers it labels 80%, and wrong on roughly 20% of them, which is what makes a threshold in your code meaningful.
TypeSafe has published no paper, no dataset description and no training detail for RLCD. The weights are not released either. As of 2026-09-18 the acronym is a name and a stated objective, with nothing an outside researcher can reproduce or measure. Take the difference from RLHF on the company’s word.
One relevant detail: TypeSafe’s team page credits CEO Diogo Almeida with co-inventing RLHF, so the contrast is being drawn by someone who worked on the thing being contrasted. RLCD explained collects what has and has not been disclosed.