the short answer
RLCD means reinforcement learning for calibrated decisions. TypeSafe AI uses the term for training models such as Jev to return bounded decisions and probabilities rather than generated text. Calibration is a property measured across groups of predictions: events assigned probability 0.8 should occur about 80% of the time. It does not guarantee that any individual 0.8 answer is correct.
- Expanded name
- Reinforcement learning for calibrated decisions
- Published framing
- TypeSafe AI’s training approach for System One models
- Optimized interface
- Decisions and probability distributions
- Important limit
- Calibration is measured over groups, not guaranteed per prediction
What RLCD Changes
RLCD changes the post-training target described by TypeSafe. RLHF trains a model toward responses humans prefer. RLVR uses verifiable rewards for tasks with checkable outcomes, commonly reasoning problems. RLCD instead targets a constrained decision and a probability distribution intended to reflect uncertainty.
The public documentation explains the objective and interface, not the complete training recipe. Claims about architecture, datasets, reward construction or independent superiority need their own primary evidence. RLCD should therefore be described as TypeSafe’s named method, not as a universally standardized research category.
Architecture: What Is Documented and What Is Not
This boundary matters when drawing architecture diagrams. The safe diagram is the observable contract: state and criteria enter, distributions leave, and external software measures outcomes. A transformer-style block diagram, parameter count or reward pipeline would be speculation unless TypeSafe publishes those details.
| Layer | Publicly documented | Not publicly specified |
|---|---|---|
| Input | Shared text state plus typed questions | Internal token representation and preprocessing details |
| Inference | A new architecture and parallel sampler; outputs are not autoregressive prose | Parameter count, layer design and serving topology |
| Post-training | RLCD targets decisions and calibrated probabilities | Reward construction, training mixture and optimization implementation |
| Output | Typed distributions; Choice/Score confidence | A generated chain of thought or explanation, because Jev does not provide one |
| Validation | TypeSafe publishes workflow evals and known jaggedness | Universal performance outside the measured tasks |
RLCD vs RLHF vs RLVR
These columns describe objectives, not a league table. An agent can use all three kinds of model: a generative model to plan and write, a reasoning model for a hard derivation, and Jev for repeated bounded judgments. The right comparison is the decision required by the software, not one model category in the abstract.
| Approach | Typical target | Typical interface |
|---|---|---|
| RLHF | Human-preferred responses | Generated text or actions |
| RLVR | Verifiably correct outcomes | Reasoning and generated solutions |
| RLCD | Calibrated bounded decisions | Typed answers and probabilities |
What Calibrated Means
Suppose a fixed question produces 1,000 predictions near 0.8. If the model is calibrated for that question and population, roughly 800 corresponding outcomes should be positive. Calibration can still coexist with mediocre accuracy, poor ranking or bad performance on a minority slice. It can also shift when the model, question wording, state distribution or language changes.
Evaluate calibration on representative labeled data. Keep a held-out set, pin the Jev model and question version, plot observed rate against predicted probability, and report a metric such as Brier score or expected calibration error alongside coverage and false-action rate. A threshold should follow the consequence of a mistake, not a generic confidence slogan.
How Software Should Use the Probabilities
- Keep the raw distribution, chosen answer, model version and question version.
- Fit thresholds on the target workflow, then verify them on held-out outcomes.
- Use separate thresholds for low-risk automation and irreversible actions.
- Route uncertain or high-consequence cases to review or a different system.
- Re-run the evaluation when an alias, model version, criteria or traffic mix changes.
FAQ
Does RLCD guarantee correct probabilities?
No. Calibration is empirical and population-specific. It must be measured on the question, data distribution and model version being deployed. A calibrated system can still make individual mistakes.
Is RLCD the same as RLHF?
No. TypeSafe describes RLHF as optimizing generated responses for human preference and RLCD as optimizing bounded decisions and calibrated probabilities. They serve different output contracts.
Is RLCD an open training recipe?
TypeSafe’s public primer explains the goal but does not document enough implementation detail to reproduce Jev. Community projects using the term are independent experiments unless TypeSafe states otherwise.
How do I test RLCD calibration?
Collect representative labeled outcomes, pin the model and question, compare predicted probabilities with observed rates, and report calibration together with accuracy, coverage and the cost of wrong automated actions.
Sources
Checked against the sources below on September 22, 2026. Model versions, prices and limits change.
- TypeSafe AI docs: AI primer and RLCD
- TypeSafe AI docs: System One
- TypeSafe AI docs: Confidence
- TypeSafe AI docs: Models
- TypeSafe AI docs: Jev 1.13 jaggedness