Decision models such as Jev answer questions with probabilities, and those probabilities are only useful if they are calibrated. A new arXiv paper notes that open-source reproductions of such models typically rely on supervised fine-tuning plus temperature scaling. The paper introduces OpenJev-RLCD, which it describes as a working RLCD implementation.
The abstract does not expand the RLCD acronym, but it clearly contrasts the approach with the SFT-plus-scaling pipeline. Where the standard approach adjusts probabilities after training, RLCD is presented as a different learning method. The authors say the implementation works, though the abstract alone does not include experimental details.
For researchers working with decision models, the value of OpenJev-RLCD is that it provides a concrete alternative to the standard calibration pipeline. If the implementation performs as described, it could be a useful baseline for comparing calibration methods. The paper is a single source, so there are no conflicting accounts to reconcile.