OmniJev/awesome-jev-gallery

OmniJev’s awesome-jev-gallery skips the usual awesome-list format of tools and demos and instead tracks the research and evaluation layer around Jev: TypeSafe’s own launch post and documented failure modes, open codebases that rebuild the System One shape, and independent evaluations of how well Jev actually performs. As of its latest listed refresh it holds 142 entries across 8 sections, including 39 open models and codebases that reproduce the typed-decision architecture and 24 independent evaluations of Jev collected separately from that reproduction work.

The list’s official section stays narrow on purpose, pointing to the load-bearing pages only: the launch post describing state-in, typed-decisions-out with RLCD training at 70 to 500 millisecond latency and $0.042 per million tokens; the three question primitives, choice, score, and noul; TypeSafe’s own admitted failure modes for Jev 1.13, including literal reading, counting, dates, and adversarial content; and the official LLM-backed adapter, which the list calls the baseline for every comparison. A recent update added 14 entries covering local and multimodal decision models, calibration tools, and comparative evaluations, and removed two older entries that had gone dead.

Open on GitHub