RLCD: Reinforcement Learning for Calibrated Decisions
RLCD is TypeSafe’s name for the training method behind Jev: Reinforcement Learning for Calibrated Decisions. The 15 September 2026 launch post contrasts it with RLHF (human preference on writeups and chat) and RLVR (verifiable rewards on problems that can be checked programmatically). This page is the mental-model explainer, not a training paper.
This is for builders who already heard of Jev / System One and need to know what “calibrated” is supposed to mean before they threshold confidence or noul in production. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Confirm naming and claims on docs.typesafe.ai. We do not sell or issue API keys.
Definition
RLCD, as TypeSafe describes it, optimizes for calibrated decisions: answers with epistemically honest probabilities on System One tasks. Jev is the first public System One model; RLCD is the named objective attached to that model. You do not turn RLCD on in a request. You consume the resulting typed fields. A short glossary lives at RLCD.
Mental model: groups, not a single verdict
Calibration is a statement about groups of predictions. If the model says 0.8 on many similar questions, about 80% of those should be correct. The System One page is careful: calibration “does not guarantee that an individual answer is correct.”
That is why the product surface emphasizes probabilities and confidence as things your code thresholds. A Noul of 0.91 on one ticket is not a certificate. A hundred Nouls near 0.9 that match your labels at about 90% is the claim you can test. Schema-safe is a different axis: Jev cannot emit a Choice outside criteria. That is not factual correctness. Schema-safe ≠ correct.
How TypeSafe contrasts RLCD with RLHF and RLVR
The launch post uses the three names as different targets:
| Name | Target TypeSafe attributes | What you should not infer |
|---|---|---|
| RLHF | Human-preferred essays, chat tone, helpful writeups | That Jev was trained to draft replies |
| RLVR | Programmatically checkable answers (math, code, puzzles) | That Jev is a calculator or solver |
| RLCD | Calibrated snap judgments on System One questions | That a public paper specifies the loss |
Those rows are vendor framing. Treat the table as a map of what they say they optimized, then measure Jev on your labels.
What TypeSafe says they did not train for
- Human-preferred essays and chat tone (RLHF’s usual target).
- Open-ended string generation. Jev is not a ChatGPT replacement.
- Per-customer fine-tunes or LoRAs. Same weights serve every account. You specialize through
state,instructions, andcriteria.
If you need a writer, compose a generative model after Jev decides. Keep arithmetic in code — jaggedness notes that jev-1.13 is weak at numeric precision.
What this site will not invent
TypeSafe has not published a full RLCD paper in the public docs index we used. We will not reconstruct loss functions, datasets, or unpublished ablations. Treat RLCD as the named training objective attached to Jev, and judge the model by the public API contract plus your own labeled traffic. Open-weight replicas are not Jev and are not RLCD — label them separately.
How this shows up in an integration
Pin jev-1.13.0 if you fit thresholds. Ask atomic questions. Log probabilities, confidence (Choice / Score), and noul. Score calibration on held-out labels — decision quality. You cannot pass rlcd: true. There is no RLCD header.
FAQ
Is RLCD a request flag or a separate model?
No. Public docs name RLCD as the training method behind Jev. You call POST /v1/systemone with jev-1.13.0 or an alias. There is no documented rlcd field.
Does “epistemically honest” mean the answer is true?
No. TypeSafe’s calibration language is about probabilities matching frequencies in aggregate. One answer can still be wrong. Schema-safe ≠ correct.
Can I reproduce RLCD from this page?
No. There is no full public paper in the sources we used. We will not invent datasets or losses. Judge the shipped model, not a reconstructed recipe.
Is OpenJev trained with RLCD?
Not as a TypeSafe product claim. Open replicas are not Jev. Label them as unofficial copies.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.