RLCD vs RLHF vs RLVR: different post-training objectives

the short answer

TypeSafe describes RLCD as reinforcement learning for calibrated decisions: training toward bounded decisions and probabilities. RLHF trains toward responses people prefer, while RLVR rewards outcomes that can be verified, often in reasoning tasks. These are different objectives, not a universal quality ranking. A production system may use an RLHF model for conversation, an RLVR model for hard reasoning and Jev for repeated typed judgments.

RLCD: Reinforcement learning for calibrated decisions
RLHF: Reinforcement learning from human feedback
RLVR: Reinforcement learning with verifiable rewards
Comparison scope: Optimization target and output contract

The Shortest Distinction Is What Receives Reward

RLHF asks people which generated response they prefer and uses that feedback to improve response behavior. RLVR uses a verifier when correctness can be checked, such as an exact answer or executable test. TypeSafe describes RLCD as training decisions and their probabilities rather than generated text. The acronyms describe optimization setups, not three interchangeable model brands.

Method	Training signal	Typical interface goal
RLHF	Human comparisons or preference-model reward	Generate responses people prefer
RLVR	A reward checked by a verifier	Produce correct outcomes on checkable tasks
RLCD	TypeSafe’s decision and probability objective	Return bounded judgments consumed by code

The Training Loops Differ at the Output Boundary

This diagram is conceptual, not a disclosure of Jev’s internal training pipeline. TypeSafe has not published the complete RLCD reward construction, data mixture, architecture or optimization recipe. The public claim is limited to the objective and decision-oriented interface.

01RLHF responseThe model generates candidate text.

02Human preferencePeople compare or rate responses; a learned reward may generalize that signal.

03RLVR solutionThe model generates an answer or reasoning trajectory.

04VerifierCode, tests or another reliable procedure checks the outcome.

05RLCD decisionThe model returns a bounded answer distribution.

06Decision feedbackTypeSafe says optimization targets decision quality and calibration.

Each method connects model output to a different source of reward.

Why This Is Not a Leaderboard

The methods optimize different interfaces. A model that writes the best explanation may be wasteful for a million binary routing decisions. A fast decision model cannot replace a model that must draft the response or solve a multi-step proof. Compare systems on the workflow outcome each one is asked to produce.

RLCD is TypeSafe’s term and public framing. The documentation does not disclose a complete reproducible Jev training recipe. Avoid inferring architecture, datasets or superiority beyond sourced measurements.

Human and Verifiable Feedback Solve Different Labeling Problems

Human preference is useful when quality is subjective or difficult to reduce to one exact checker, but reviewers can disagree and reward style instead of factual success. Verifiable rewards are powerful when a test, proof checker or exact answer faithfully represents success, but many social and product judgments do not have such a verifier.

A calibrated decision objective is useful when software needs probabilities over predefined outcomes. It still requires labels or outcomes that represent the intended decision. Calibration cannot repair a vague question, incomplete state or wrong target.

Evaluate the Objective You Need

For conversation, measure task success, preference and safety of generated responses.
For reasoning, measure verified solution correctness and resource use.
For decisions, measure class errors, calibration, coverage, latency and downstream action cost.
For a hybrid system, measure the complete routed workflow and every fallback path.

A Production Agent Can Use Models Trained with Different Objectives

The model used for each step should be selected from measured task performance, not the acronym attached to its training. An RLHF-trained model can still classify; an RLVR-trained model can still generate; Jev can still fail a bounded judgment. The interfaces and validation evidence determine where each is safe to use.

System step	Suitable capability	Example
Generate	Instruction-following language model	Draft the support reply
Reason and verify	Reasoning model plus deterministic verifier	Produce code and run tests
Judge	Jev bounded decision	Estimate whether the reply claims unsupported completion
Authorize	Application rules and identity system	Permit or reject a refund action
Escalate	Human review	Resolve ambiguous or consequential exceptions

Keep Evidence Claims at the Right Level

Cite TypeSafe when defining RLCD; it is the company’s term and framing.
Do not infer Jev’s parameter count, layers, data mixture or exact reward model.
Do not claim RLCD guarantees calibration on an unseen application.
Cite the original paper or technical report when describing RLHF or RLVR examples.
Compare deployed systems on shared labels rather than ranking training acronyms.

FAQ

Is RLCD better than RLHF?

Not in the abstract. RLCD targets bounded calibrated decisions, while RLHF is suited to preferred generated responses. The task determines the useful objective.

Is RLCD the same as RLVR?

No. RLVR rewards verifiably correct outcomes, commonly for reasoning. TypeSafe describes RLCD as optimizing decisions and probabilities for calibration.

Does Jev use RLCD?

Yes. TypeSafe’s public model documentation says Jev is trained with RLCD. The complete training implementation is not publicly documented.

Can one application use all three?

Yes. A system can use different models for conversation, reasoning and bounded decisions while code routes work and enforces permissions.

Sources

Checked against the sources below on September 22, 2026. Model versions, prices and limits change.