RLCD: Reinforcement Learning for Calibrated Decisions
RLCD is TypeSafe’s name for the training method behind Jev: Reinforcement Learning for Calibrated Decisions. The launch post contrasts it with RLHF (human preference on writeups and chat) and RLVR (verifiable rewards on problems that can be checked programmatically).
RLCD, as TypeSafe describes it, optimizes for calibrated decisions: answers with epistemically honest probabilities on System One tasks.
What that is supposed to mean in product terms
Calibration is a statement about groups of predictions. If the model says 0.8 on many similar questions, about 80% of those should be correct. The System One page is careful: calibration “does not guarantee that an individual answer is correct.”
That is why the product surface emphasizes probabilities and confidence as things your code thresholds, rather than a single “the model said so” bit.
What TypeSafe says they did not train for
- Human-preferred essays and chat tone (RLHF’s usual target).
- Open-ended string generation.
- Per-customer fine-tunes or LoRAs. The models page says Jev is not fine-tuned with customer data; the same weights serve every account. You specialize it through
state,instructions, andcriteria.
What this site will not invent
TypeSafe has not published a full RLCD paper in the public docs index we used for this draft. We will not reconstruct loss functions, datasets, or unpublished ablations. When they publish more, update this page from that source.
Until then, treat RLCD as the named training objective attached to Jev, and judge the model by the public API contract plus your own labeled traffic.
Sources
Public TypeSafe or adjacent documentation only. No private claims.