Glossary: RLCD
RLCD means Reinforcement Learning for Calibrated Decisions. It is TypeSafe’s name for how Jev (the first public System One decision model) is trained to emit epistemically honest probabilities on Choice, Score, and Noul questions. TypeSafe contrasts RLCD with RLHF (human preference on writeups and chat) and RLVR (verifiable rewards on problems that can be checked programmatically) in the launch post. That contrast, and the “honest probabilities” phrasing, is a vendor claim. Public docs name the method; they do not publish a full paper, dataset card, or loss schedule.
This glossary entry is for builders who already heard of Jev / System One and need a citable definition before they treat confidence or noul as a warranty. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Confirm naming against docs.typesafe.ai. We do not sell or issue API keys.
How it shows up
You cannot enable RLCD as a request flag on POST /v1/systemone. You consume the resulting probabilities, confidence (Choice and Score only), and noul values, then measure calibration on your labels. The models page says Jev is not fine-tuned with customer data; the same weights serve every account. You specialize the model through state, instructions, and criteria, not by turning on a training mode.
Open-weight replicas are not RLCD and are not Jev. Label them as unofficial clones. Pin a versioned model id such as jev-1.13.0 before you compare weeks: jev-latest can move when TypeSafe ships, and that movement is not a change you requested in the training recipe.
Contrast
RLCD vs calibration. RLCD is the named training objective. Calibration is the property you can measure: events with reported p ≈ 0.8 should happen about 80% of the time across a group. TypeSafe says RLCD trains for that property. You still plot ECE or a reliability diagram on your traffic.
RLCD vs confidence. Confidence is a 0–1 summary of how peaked a Choice or Score probabilities map is. It is a derived statistic on the wire, not a training hyperparameter you send. Noul has no confidence field; do not invent one because “RLCD implies confidence.”
RLCD vs RLHF / chat fine-tunes. RLHF optimizes for preferred text. Jev does not generate replies, code, or reasoning essays. Asking Jev to “write a safer answer” is the wrong product. Ask a typed question; let code act.
RLCD vs System One. System One is the product class (fast structured decisions). RLCD is TypeSafe’s name for the training story attached to that class’s first public instance, Jev.
Common mistakes
- Inventing a loss function, reward model, or unpublished ablation and citing jev.pro as if it were the paper.
- Treating launch-post plots as your ECE. Vendor plots are not your production reliability diagram.
- Assuming OpenJev or other replicas inherit RLCD. They do not.
- Reading high confidence as a per-row guarantee. Calibration is group-level.
- Sending
rlcd: trueor a similar invented field. Schema-safe requests still fail 422 if you add unknown required shapes — and even a 200 does not prove the winner is factually correct.
See also
FAQ
Does jev.pro reproduce RLCD?
No. We cite TypeSafe’s naming. We do not invent a training recipe, loss, or dataset.
Is OpenJev trained with RLCD?
No. Label open-weight clones as not Jev and not endorsed.
Can I request an RLCD-trained custom model?
Public docs we used say the same weights serve every account. Specialize with state, instructions, and criteria. Do not assume a per-customer LoRA.
Does RLCD apply to Noul?
TypeSafe describes calibrated decisions across System One tasks. Noul still has no confidence field; calibrate the noul itself.
Limits
This is a jev.pro working definition unless TypeSafe uses the same word as an API field. We do not invent search volume or unpublished math. Schema-safe answers are not automatically factually correct. When TypeSafe publishes a paper, update this page from that source.
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.