RLCD Explained: Reinforcement Learning for Calibrated Decisions
RLCD (Reinforcement Learning for Calibrated Decisions) is the training method TypeSafe AI says it used for Jev, its System One model. Where RLHF (Reinforcement Learning from Human Feedback) optimizes a chat model for responses humans rate as good, and RLVR (Reinforcement Learning from Verifiable Rewards) optimizes for tasks with a checkable right answer (like a math proof or a passing test), RLCD is described as optimizing specifically for honest, calibrated probabilities on structured decision tasks.
Quick answer: RLCD trains a model so its confidence scores mean what they say — a 0.7 should be right about 70% of the time — rather than training it to produce responses humans prefer (RLHF) or to solve tasks with a verifiable ground-truth answer (RLVR). TypeSafe AI states Jev was trained this way; the mechanics beyond that framing haven't been published in detail as of launch, so treat the description as vendor-stated positioning rather than a fully specified, independently verified method.
What problem does RLHF solve, and why isn't it the right fit here?
RLHF trains a model to produce outputs that human raters prefer — it's how chat models learn to be helpful, follow instructions, and avoid unwanted content. The reward signal is "did a human like this response," which is great for open-ended generation but doesn't directly optimize for whether a stated confidence number is statistically accurate. A model can learn to sound confident (because raters like confident-sounding answers) without its confidence tracking reality at all.
What problem does RLVR solve, and why isn't it the right fit either?
RLVR trains on tasks with a verifiable, checkable answer — the reward comes from an automated checker (does the code pass tests, is the math proof valid), not a human rating. It's well suited to tasks like coding and math where correctness is unambiguous. But most of Jev's target tasks — routing, scoring, risk classification — don't have a single objectively "verifiable" answer the way a math proof does; they have a distribution of plausible outcomes, and what you actually want is an honest probability over that distribution, not a single verified answer.
What does RLCD optimize for instead?
Per TypeSafe AI's positioning, RLCD optimizes directly for calibration on structured tasks — reward the model not just for picking a plausible category or score, but for its stated probability matching the actual observed frequency of outcomes. This is the training rationale behind Jev's Score and Noul question types in particular, where the entire value of the output depends on the number being calibrated, not just directionally correct. See calibrated confidence scores for what calibration means and how to check it independently, regardless of which training method produced the score.
| Method | Optimizes for | Typical use |
|---|---|---|
| RLHF | Responses humans rate as good | Chat models, general assistants |
| RLVR | Verifiable, checkable correctness | Math, code, tasks with a ground-truth checker |
| RLCD | Calibrated probability of an outcome | Structured decisions: classification, scoring, yes/no (Jev) |
Does RLCD training explain Jev's accuracy numbers?
Not directly — calibration and accuracy are different properties. A model can be well-calibrated (its 0.7s are right 70% of the time) without being highly accurate (most of its predictions being correct), and vice versa. TypeSafe AI reports Jev at 67.8% accuracy on its own four-workflow benchmark, versus 67.9% for GPT-5.6 Terra, 74.1% for Sol, and 73.1% for Opus 5 — these are vendor-reported and not independently reproduced. RLCD's claimed benefit is that when Jev is wrong, its confidence should still be honest about how likely that is, rather than that RLCD makes Jev more often right. Whether that holds up requires independent calibration testing on real workloads, which hasn't happened yet at five days post-launch.
Common pitfalls
A few mistakes come up repeatedly when reasoning about RLCD:
- Assuming RLCD is a drop-in replacement for RLHF or RLVR in general — it's aimed specifically at calibrated structured decisions, not chat quality or verifiable-task correctness.
- Treating "trained for calibration" as proof of calibration — verify it yourself with a reliability diagram on your own data.
- Confusing calibration training with accuracy improvement — they're different objectives with different evidence needed.
- Repeating TypeSafe's description of RLCD as an established, peer-reviewed method — as of launch it's vendor terminology without published methodology detail we've been given.
Pro Tip: When evaluating any model trained for calibration — RLCD or otherwise — ask for (or build yourself) a reliability diagram, not just an accuracy number. Accuracy tells you how often it's right; calibration tells you whether you can trust its stated confidence when it's wrong.
← Back to AI & ML Tips