v1.0.0

Typed decisions (yes/no, choice, score) on any JSON or text, inside a workflow, on the runner CPU. No API key, nothing leaves the runner.

• workflow: one of four built-ins (security_incidents, agent_trace_observability, customer_service, invoice_processing) is answered by the ensemble: 0.6970 on the typed-decisions test split.
• questions: your own typed questions run on the fine-tuned model alone: 0.6235, and unmeasured on questions unlike the training set. Check outputs before automating.

Model files come from the model-v1 release and are checked against action/SHA256SUMS. First run downloads ~1.2 GB, then caches. Use @v1.

Model v1

Model files for Open Jev and its GitHub Action.

Checkpoint (PyTorch, fp16, 298 MB) - uploaded in six pieces:

gh release download model-v1 -R intikhab49/open-jev-typed-decision-engine -p 'jevlite-fp16.pt*'
cat jevlite-fp16.pt.part-* > jevlite-fp16.pt
sha256sum -c jevlite-fp16.pt.sha256

Load it like jevlite.pt, casting the floating weights back to fp32 (v.float()), as action/build_assets.py does. Measured fp16 round trip on 200 test decisions: same answer on 200/200, largest probability shift 0.0006.

The ONNX files the Action downloads are built from this checkpoint by the build-model workflow and added to this release.