The short version
Three acronyms now describe how frontier models are shaped. RLHF rewards answers human raters prefer. RLVR rewards answers a program can verify. RLCD — TypeSafe's method — rewards epistemically honest probabilities on decision tasks.
The first two produce excellent text. The third produces something different: a model whose confidence means something.
Why this matters
If a model can do a task correctly 95% of the time but cannot tell you which 5% it is unsure about, you cannot safely automate that task. You either accept the errors or you keep a human on every case.
Calibration is what makes the threshold pattern possible: act above a confidence line, escalate below it. That only works if the score is trained to mean something rather than elicited by a polite request.
What each method optimises for
It helps to be blunt about the objective, because the objective is the behaviour.
- RLHF: human raters pick the response they prefer — great for chat, prone to mode dropping and overconfidence
- RLVR: rewards can be checked programmatically — great for maths and code, useless when there is no ground truth to check
- RLCD: rewards calibrated probabilities on decisions — great for automation, worthless for writing prose
- Consequence: prompting an RLHF model for a confidence score gives you a number, not a calibrated one
The architecture that comes with it
TypeSafe pairs RLCD with a new architecture and a parallel sampler, which is where the 70-500ms latency band comes from. Training for calibration and generating all outputs in one pass are two halves of the same design decision: decisions, not strings.
That is also why type errors are impossible: the schema is fixed before inference, so there is nothing to hallucinate. See Jev vs LLMs.
Common mistakes
Vendors and buyers both misread the training difference.
- Assuming RLCD makes a model smarter — it makes it more honest about what it knows
- Bolting confidence prompts onto RLHF models and calling the output calibrated
- Comparing models on benchmarks that were never the objective
- Expecting RLCD models to explain themselves; there is no rationale in the output
How PixaSocial Ai helps
PixaSocial Ai keeps audience and brand context in structured form, which is what makes calibrated decisions actionable — a probability is only useful when it is attached to a clearly defined judgement. More in the articles hub.