RLCD: how Jev is trained

RLCD stands for Reinforcement Learning for Calibrated Decisions, the training method behind Jev. Instead of optimising for fluent writing, it optimises for calibration — so when Jev says 80%, it is right about 80% of the time.

Most large models are trained to produce convincing text. RLCD takes a different aim: it trains Jev so that its stated probabilities line up with real-world outcomes. That single shift in objective is what separates a calibration-trained decision model from a general chatbot.

What RLCD optimises for

The training objective of RLCD is calibration, not eloquence. The model is rewarded when its confidence matches how often it is actually correct: if it says an answer is 80% likely, that answer should hold up roughly 80% of the time across many cases. Fluent prose is not the goal — well-calibrated numbers are.

How it shows up in every answer

Because Jev is trained with RLCD, every response carries more than a verdict. Each answer includes:

• A probability distribution over the possible outcomes.
• A confidence value describing how concentrated that distribution is.
• A typed result locked to the shape you defined.

That RLCD calibration is what makes those probabilities trustworthy enough to build on. Without it, a confidence number would be decoration; with it, it is a signal you can threshold against.

Zero hallucination versus calibration

It helps to separate two ideas. Zero hallucination means Jev's output is locked to the type you defined, so it cannot return an option you never listed or a malformed shape — that guarantee comes from the typed contract, not from RLCD. Calibration is the separate promise the method delivers: confidence tells you how concentrated an answer is, not whether it is correct.

Using RLCD confidence in production

Treat the confidence as a dial, not a certificate. Calibrate your own thresholds on historical labelled data, then route high-confidence decisions automatically and send low-confidence ones to review. Used this way, RLCD lets you trade coverage for accuracy on purpose rather than by guesswork — the practical payoff of Reinforcement Learning for Calibrated Decisions.

See also: What is Jev · Use cases · Playground

Related: Jev architecture · System One model · Jev vs an LLM