Qualification: TypeSafe, Introducing System One models and Jev (TypeSafe blog, 2026), and the promotional claim that Jev is what comes after RLHF

Record: kaal:position:2026-09-21-009 · 2026-09-21

Calibration can only be measured against a reference. Where no answer key exists, that reference is a label someone supplies, whether a human annotator, another model or a synthetic generator. If that label is costless, the flaw RLHF had returns through the label.

Affirmed commentary position. This record extends a source-bound scholarly claim but is not a verbatim paper claim.

Holds when

TypeSafe's post announces a product. It is not an empirical test of these claims.
The position is Kaal's own argument in the essay After RLHF Comes a Stack, Not a Model.
The position does not evaluate Jev's accuracy, which TypeSafe has not published a calibration figure for.
The accountable-validation layer is proposed, not tested: no policy has yet been trained under it.

Current debate

TypeSafe, Introducing System One models and Jev (TypeSafe blog, 2026), and the promotional claim that Jev is what comes after RLHF

Scholarly basis

kaal:claim:7456999-024
Wulf A. Kaal, Reinforcement Learning from Reputation Feedback (2026). SSRN: https://ssrn.com/abstract=7456999
Source PDF sha256: 01016fbbcc25cad205cebd59ea686d8dfee8ae9b632095d86ba6c9a900456a4b

Topics

ai-and-agents research-methods risk-and-incentives

Provenance

Affirmed in kaalvis:2026-09-21:rlhf-stack-essay-six-v1 on 2026-09-21. Review record.

Verify

Canonical markdown sha256: fca431ad631d410b864ae87d6c9dcf6fe34d577f5eceb5da3887735c5fa86de2
curl -s https://wulfkaal.github.io/positions/2026-09-21-009.md | sha256sum