RLCD (2026) | Reinforcement Learning for Calibrated Decisions, TypeSafe AI Jev, Brier Score, ECE & Calibration Resources

What Is RLCD?

3 entries

Reinforcement Learning for Calibrated Decisions (RLCD) is the name TypeSafe AI gives to the training method behind Jev, its first “System One” model, announced in September 2026. TypeSafe describes Jev as returning typed, structured decisions with calibrated probabilities rather than generated text, and contrasts RLCD with RLHF, which optimizes for human preference, and RLVR, which rewards verifiable correctness. Calibration means that a model's stated confidence matches how often it is right: among predictions made with 80 percent confidence, about 80 percent should be correct. The ideas commonly used to measure and train for it are well established: expected calibration error (ECE), the Brier score and other proper scoring rules, and reinforcement learning methods such as PPO and GRPO. TypeSafe has not published full technical details of RLCD, so third-party descriptions, including the article this page follows, are interpretations rather than specifications. The article is by Utkarsh Ankit in Towards AI (September 2026), and says its reward design is the author's reconstruction from standard calibration methods; the text here is written independently. Performance, price and accuracy figures for Jev are the company's own claims.

01Source articleRLCD: Reinforcement Learning for Calibrated Decisions, by Utkarsh Ankit, Towards AI. 02Introducing System One models and JevTypeSafe AI's announcement, which names RLCD and compares it with RLHF and RLVR. 03JEV-as-a-JudgeCarnegie Mellon paper evaluating Jev as a confidence-gated judge, in a cascade that escalates uncertain cases.

TypeSafe AI & Jev

11 entries

Primary sources and coverage.

04TypeSafe AICompany behind Jev and RLCD. 05TypeSafe AI blogAnnouncements and essays from the company. 06The Bitterest LessonTypeSafe AI essay on its approach. 07TypeSafe documentationAPI and schema documentation for Jev. 08Jev explained (Sanity glossary)Glossary entry on Jev; third-party. 09Building a harness with Jev (LangChain)LangChain's guide to Jev; third-party. 10How to use Jev (DEV Community)Practical guide; third-party. 11Jev and RLCD tutorial (Analytics Made Simple)Tutorial on Jev and RLCD training; third-party. 12RLCD explained (OrcaRouter)Explainer; third-party. 13Jev explained (DataCamp)DataCamp explainer; the site blocks automated checks. 14MarkTechPost reportNews report on the release; the site blocks automated checks.

Calibration Foundations

13 entries

Measuring whether confidence matches accuracy.

15On Calibration of Modern Neural NetworksGuo et al. (ICML 2017): shows modern networks are often overconfident and popularized ECE and temperature scaling. 16Calibration (statistics)Overview of calibration in statistics. 17Brier scoreMean squared error between predicted probabilities and outcomes, and its decomposition. 18Scoring ruleProper scoring rules, which are optimized in expectation by reporting true probabilities. 19Brier (1950), Monthly Weather ReviewThe original paper on verification of probability forecasts. 20scikit-learn: probability calibrationReliability curves, Platt scaling and isotonic regression. 21scikit-learn: brier_score_lossComputing the Brier score. 22scikit-learn: IsotonicRegressionNon-parametric recalibration. 23TorchMetrics: calibration errorECE metric for PyTorch. 24Temperature scaling reference codePyTorch implementation accompanying Guo et al. 25Calibration framework (netcal)Python library of calibration methods and metrics. 26Focal lossLoss function sometimes used to reduce overconfidence. 27Label smoothing (Rethinking Inception)Origin of label smoothing, a training-time regularizer.

LLM Confidence & Uncertainty Research

5 entries

How language models can report calibrated confidence.

28Beyond Binary Rewards (RLCR)Trains language models to reason about their uncertainty by adding a calibration term to a verifiable reward. 29Language Models (Mostly) Know What They KnowAnthropic study of how well models predict whether their own answers are correct. 30Just Ask for CalibrationStrategies for eliciting calibrated confidence from RLHF-tuned models. 31Can LLMs Express Their Uncertainty?Empirical evaluation of confidence elicitation methods. 32Generating with ConfidenceUncertainty quantification for black-box language models.

Reinforcement Learning Methods & Tooling

8 entries

The training machinery calibrated rewards plug into.

33PPOProximal Policy Optimization. 34DeepSeekMath (introduces GRPO)Paper that introduced Group Relative Policy Optimization. 35TRL GRPO trainerHugging Face documentation for GRPO training. 36TRLHugging Face library for post-training language models with RL. 37verlReinforcement learning library for LLM post-training. 38OpenRLHFOpen-source RLHF framework. 39OpenAI Spinning UpEducational resource on deep reinforcement learning. 40Stable-Baselines3Reliable reinforcement learning implementations in PyTorch.

9 entries

Neighbouring pages on this site.

41Reinforcement LearningReinforcement learning hub. 42RL AlgorithmsReinforcement learning algorithms. 43Reinforcement Learning GuideGuide to reinforcement learning. 44Probability Theory for AIProbability foundations for AI. 45Hallucination Mitigation MechanismsReducing LLM hallucination. 46LLM Training and Inference OptimizationTraining and inference techniques. 47LLMOpsOperating LLM applications. 48LLM HubLarge language model hub. 49Machine LearningMachine learning hub.

Frequently Asked Questions

What does RLCD stand for?

Reinforcement Learning for Calibrated Decisions. It is the name TypeSafe AI gives to the training method used for its Jev model.

What is calibration?

A model is calibrated when its stated confidence matches how often it is correct. If it says 80 percent for many predictions, about 80 percent of them should turn out right. Calibration is related to, but different from, accuracy and raw confidence.

What are ECE and the Brier score?

Expected calibration error groups predictions into confidence bins and averages the gap between confidence and accuracy in each bin. The Brier score is the mean squared error between predicted probabilities and actual outcomes, and it is a proper scoring rule, so reporting true probabilities is optimal.

How does RLCD differ from RLVR?

RLVR typically rewards a correct answer with 1 and an incorrect one with 0, which gives no direct reason to report honest confidence. TypeSafe says RLCD instead optimizes for calibrated probabilities. Full details are not public, so descriptions of the exact reward are third-party interpretations.

Does Jev never hallucinate?

TypeSafe says Jev cannot hallucinate because it returns typed structured values rather than free text. Commentators note that this concerns format validity, not whether an answer is correct, and the company's performance figures are its own.

How current is this page?

Links were checked in October 2026. RLCD and Jev are new and TypeSafe has not published full technical details, so performance and price figures come from the company and should be treated as vendor claims. Sites that block automated checks are kept and may need a manual click.