OpenJev v0.3.0 — experimental shared-state decisions

First public experimental release of OpenJev: local typed probabilistic decisions with Qwen3-0.6B, a shared state prefix, isolated candidate branches and a learned scalar scoring head.

Included

• Reproducible data preparation, training, development-only checkpoint selection, calibration and evaluation.
• English and Chinese documentation, model card, source attribution and contribution guidance.
• CPU CI, cache output/gradient equivalence checks and safe checkpoint downloads.
• Historical v0.1/v0.2 baselines and complete measured regressions.

Quick start

git clone https://github.com/IamBusy/OpenJev.git
cd OpenJev
uv sync --frozen --extra qwen --extra dev
uv run --no-sync openjev-branch download
uv run --no-sync openjev-branch predict --input examples/refund.json

The model ZIP contains the LoRA adapter, scalar head, calibration and provenance. Qwen base weights are downloaded separately from a pinned Hugging Face revision. No provider key is required for inference or frozen v0.3 data reconstruction. SHA256SUMS covers the attached model ZIP, wheel and source distribution.

Evidence and limitations

• 31 local checks passed; source-only tests: 24 passed and 7 explicit integration skips. Public Linux CI passed.
• All reconstructed public split hashes and the complete v0.3 manifest match the original experiment.
• Fixed 60-question subset: v0.3 45/60, v0.2 39/60, DeepSeek 59/60.
• Long synthetic state / 8 questions × 4 candidates: about 0.70 s versus v0.2's 1.38 s on the measured M3 Pro. Short single-question inputs are slower (104 ms versus 55 ms).
• Some banking, access-policy and shipping tasks regress. Calibration is domain-specific. This is one seed with narrow data; architecture and data changed together.

OpenJev is independent of TypeSafe. It does not reproduce Jev's unpublished weights, architecture or RLCD and is not production-ready.

Full results · Model card · Reproduction guide