After RLHF Comes a Stack, Not a Model

Calibration is not accountability. The debate about what comes after RLHF keeps confusing the two.

The claim

A promotional post circulating this week presents TypeSafe’s Jev as what comes after RLHF. The post says “the next era is not the Claude Code era.” It says “Jev is the next step.” It says “This is what comes after RLHF.” It says “every model you use has a human-shaped flaw baked into it.” The company’s founder, Diogo Almeida, is a former OpenAI researcher and an author of the InstructGPT work that laid the groundwork for ChatGPT (The Decoder). Jev is trained with a method TypeSafe calls Reinforcement Learning for Calibrated Decisions, or RLCD (TypeSafe). The post’s argument is that RLCD removes the flaw.

The diagnosis is half right. The prescription is not.

Half right

RLHF has a structural flaw. I have made that argument myself (Kaal 2026b). But the flaw is not that humans sit in the loop. The paper’s position is that the flaw is institutional: whoever supplies the reward signal bears no consequence for being wrong. An annotator who mislabels a comparison loses nothing. A model that labels in the annotator’s place loses nothing either. The signal is costless, so it is cheap to corrupt and impossible to hold to account.

That is an institutional flaw, not a biological one. Replacing the human does not fix it. Giving the signal a price does.

What Jev is

Jev is a decision engine. A developer defines a question and a closed set of answers. The model scores the options and returns a probability. It does not write text, code or email. That is a useful class of system. Much of the work inside software is exactly this: classify, route, triage, gate.

It is a classification, not a successor. RLCD is a training objective for one class of outputs. RLHF is a way of sourcing the signal that tells a model what good output is. The two answer different questions. Calibration answers: when the model says ninety percent, is it right ninety percent of the time? RLHF answers, badly: who decides what right means? Calibration presupposes an answer to the second question. It does not supply one.

Three reasons Jev alone cannot come after RLHF

First, calibration needs a key. Calibration can only be measured against a reference. Where the outcome is checkable, that reference can be the outcome itself: the payment cleared or it failed, the customer left or stayed. RLCD can be trained against those outcomes without a human label. TypeSafe already sets that case apart as RLVR, so keyed territory is not the claim Jev is making. Where no key exists, the reference is a label that someone supplies. Is this answer fair? Is this clause enforceable? Is this post misleading? Someone has to say. That someone is a human annotator, another model, or a synthetic generator. If that label is costless, the human-shaped flaw returns through the label. TypeSafe has not published its loss, its reward, or a calibration figure. Its published workflow tests use other models as the reference, not independently verified outcomes (The Decoder).

Second, zero hallucination is a redefinition, not an elimination. A model that must pick from a closed list cannot invent a citation. It can still pick the wrong item on the list. Even sympathetic coverage makes the point: the model will not answer outside the preset options, but a factually wrong choice within them remains possible (The Decoder). A perfectly calibrated model that reports ninety percent is wrong about one time in ten among the decisions it reports at that rate. That is the virtue of calibration: it prices the error. It does not remove it. The error moves from fabricated text to confident misclassification. For automation, the operative question becomes who answers for the tenth case.

Third, calibration is a property of populations, not of decisions. A model can be calibrated on average and miscalibrated on a subgroup. It can drift when the world shifts. It can look calibrated on a test set that leaked into its training data. Someone has to keep checking it, on cases it has never seen. And that someone must bear a cost for checking carelessly, or the check decays into the same costless signal RLHF produced.

What comes after RLHF

No single model comes after RLHF. A stack does.

1. Verifiable rewards where an answer key exists. Mathematics, code that must pass tests, outcomes that resolve.
2. Calibrated decisions for typed choices. Engines like Jev belong here. They turn a judgment into a number software can act on.
3. Accountable validation where no key exists. This is the accountable signal RLHF never produced. The paper proposes validators who stake reputation that cannot be bought or transferred. The proposed update makes errors cost more than agreement earns. The measured quantity is a record earned where the work is checked, not the pool weight the mechanism specifies.

The paper proposes that third layer as Reinforcement Learning from Reputation Feedback (Kaal 2026b). It does not propose the stack as a tested system. The measured evidence so far is narrow but pointed. Across thirty-six parties with repeated measurement, a record earned on checked cases predicted accuracy on withheld cases at a rank correlation of +0.781. The paper notes that the effective number of independent units is closer to twenty. A party that failed the checked cases passed the withheld ones in zero of 168 observations. No policy has yet been trained under the mechanism. That is the next step, and the paper says so.

And the generative layer does not disappear. Jev does not write. The code, the drafts and the arguments still come from generative models. The Claude Code era is not ending. It is acquiring a decision layer beneath it.

Why the distinction matters

Computative Economics starts from one observation: under computational abundance, the binding constraint migrates from generation to evaluation and alignment (Kaal 2026a). TypeSafe’s published claim is cheaper typed decisions. That does not supply judgment where no key exists, and that is precisely where RLHF failed.

The successor to RLHF will not be a model. It will be the signal someone is willing to stand behind. A score is not a stake. Only consequence prices a judgment.

References

Kaal, Wulf A. 2026a. “Computative Economics: A Framework for Economic Analysis under Computational Abundance.” SSRN. https://ssrn.com/abstract=7261481.

Kaal, Wulf A. 2026b. “Reinforcement Learning from Reputation Feedback: An Alignment Mechanism Grounded in Computative Economics.” SSRN. https://ssrn.com/abstract=7456999.