JEVANY / DOCUMENTATION

Acknowledgements

JevAny includes modified infrastructure code from Kev, created by Jared Palmer and released under Apache-2.0. The required attribution is also recorded in NOTICE and in derived source files.

JevAny develops a separate research direction around calibration-aware reinforcement learning for decision models. Tianxin Wei created the current training system, RLCR implementation, checkpoints, evaluation, release tooling, and product roadmap.

The architecture was informed by the public analysis in Jev's Architecture Unmasked and the public System One API.

RLCR follows the reward proposed in Beyond Binary Rewards: Training LMs to Reason about Their Uncertainty by Damani et al. JevAny adapts it to a prefill-only pointer model.

Qwen3.8-27B is produced by the Qwen team. The adapters in this project require its separately distributed base weights.

The Phi-4 vision conversion helper adapts configuration and weight-name mappings from Hugging Face Transformers under Apache-2.0. It merges Microsoft's pretrained vision LoRA before JevAny training.

The native Phi-4 Reasoning Vision adapter follows Microsoft's published model layout and image preprocessing, released under MIT. Its weights are downloaded separately from Microsoft.

The README teaser uses Lucide icons. The source records and license notices accompany the editable SVG.

The supported-model cards use logos from Lobe Icons under MIT. Their source records and license accompany the SVG. Model and publisher marks belong to their respective owners.