What is RLCD?

RLCD is TypeSafe's name for Reinforcement Learning for Calibrated Decisions. People searching RLCD AI, Jev RLCD, or just RLCD after the Jev launch usually want this, not a generic RL paper.

RLCD vs RLHF, in TypeSafe's table

The launch post compares existing LLMs with System One + Jev:

That is TypeSafe's framing. This hub does not have an independent training-run writeup.

What “calibrated” means for you

Official confidence docs: calibration is measured across groups of predictions. Higher confidence should mean higher accuracy in aggregate. It does not guarantee that one answer is correct.

Practical use:

• Read confidence on Choice and Score.
• Read noul on Noul (there is no separate confidence field).
• Threshold in code. Escalate when the distribution is flat.

Where RLCD sits in the stack

architecture + parallel sampler + RLCD
        │
        ▼
     Jev (System One model)
        │
        ▼
  Choice / Score / Noul + probabilities

See How Jev works for the request loop and parallel sampler, and Is Jev an LLM?.

When this page is enough

You needed the acronym expanded and a sourced contrast with RLHF.

When it is not

If you need a reproducible training recipe, weights, or a paper PDF: TypeSafe has not published those on the pages we verified. Do not treat this guide as a methods paper.

Common mistakes

• Using “RLCD” as proof that a single noul of 0.91 is true.
• Confusing RLCD with RLHF because both start with “reinforcement learning”.
• Searching only “RLCD” and landing on unrelated contrastive-distillation papers. In the Jev launch window, the intended expansion is TypeSafe's.

Next: System One or How Jev works.