RLCD: Reinforcement Learning for Calibrated Decisions
Last checked · Independent guide, not affiliated with TypeSafe AI
RLCD, short for Reinforcement Learning for Calibrated Decisions, is the post-training method TypeSafe AI uses to build Jev. Instead of rewarding text that people prefer (RLHF) or answers a checker can verify (RLVR), it trains the model to return decisions whose probabilities match how often those decisions turn out to be right.
This page is about the AI training method. If you searched for RLCD screens, those are reflective LCD displays, which have nothing to do with Jev.
Three ways to post-train a language model
Section titled “Three ways to post-train a language model”TypeSafe frames RLCD as the third of three approaches for turning a pretrained language model into something useful:
| Method | Rewards the model for | Produced | Trade-off |
|---|---|---|---|
| RLHF (reinforcement learning from human feedback) | Responses people prefer | Chatbots such as ChatGPT | Can reward flattery and confident-sounding mistakes |
| RLVR (reinforcement learning with verifiable rewards) | Answers a program can check | Reasoning models, strong at math | Slower and more expensive |
| RLCD (reinforcement learning for calibrated decisions) | Decisions with honest probabilities | Jev | Gives up free-form text |
TypeSafe’s CEO, Diogo Almeida, co-invented RLHF while at OpenAI. Part of TypeSafe’s pitch is that the method he helped create is the wrong objective for automation.
What “calibrated” means
Section titled “What “calibrated” means”A model is calibrated when its probabilities match reality across many predictions. If Jev answers a thousand questions with a probability of 0.8, about 800 of those answers should be right. Answers at 0.2 should be right about 20% of the time.
Two things follow:
- Calibration is a property of groups, not of one answer. A single answer at 0.95 can still be wrong. TypeSafe’s docs say this explicitly.
- It makes thresholds meaningful. If probabilities are honest, you can decide in code that anything below 0.5 goes to a person and anything above 0.9 is automated, and predict roughly how many mistakes each band will contain.
Why TypeSafe says RLHF is not enough
Section titled “Why TypeSafe says RLHF is not enough”TypeSafe’s primer argues that preference training has two side effects that matter for unattended software:
- It rewards sounding right. People prefer confident, agreeable answers, so the model learns to produce them even when it should be unsure.
- It narrows the output distribution. TypeSafe calls this mode dropping, a milder cousin of the mode collapse seen in generative adversarial networks: the model concentrates on a favored style and loses probability mass on alternatives.
For a chatbot, those are acceptable costs. For a system that acts without a human reading the output, TypeSafe’s position is that honest uncertainty matters more than a pleasing answer.
What has not been published
Section titled “What has not been published”As of September 19, 2026, TypeSafe has released no paper, training details or calibration curves for RLCD. On Hacker News, commenters asked what exactly changes in the training objective, how it differs from RLVR in practice, and whether anything is published on how calibration is maintained. Those questions remain open. The launch post offers evaluation results but describes the method only at a high level.
Using calibrated probabilities well
Section titled “Using calibrated probabilities well”Calibration only helps if the question is well posed. In our own test, we asked Jev two questions about the same message, “The second charge on my card looks wrong to me. What happens now?”:
- “Is the customer asking for money back?” returned 0.50, genuinely undecided, because the customer never says it.
- “Does the customer say a charge might be wrong?” returned 0.97.
The model was not being vague in the first case. The question was. See Noul and Confidence for how to write questions and pick thresholds.