Open Jev — JoshuaSP

No. It is an independent experiment using an open diffusion model, not a reproduction of Jev architecture, training, calibrated probabilities, or published workflow scores.

Grouping questions that share a document into one canvas cut inference time and GPU cost sharply with a small accuracy drop. Oversized nodes split to fit the token limit and merge back into complete objects.

The published numbers all come from Modal: Python 3.11 or newer, a Modal account, and an HF token whose DiffusionGemma access terms you accepted, stored as the Modal secret huggingface. Each run is a bounded H100 job rather than a persistent endpoint, with weights cached in the open-jev-hf-cache volume. verify.py checks tokenizer and constraints on CPU with no weights at all. For local inference you install the inference extra and supply your own CUDA VRAM. Bitwise reproduction is not promised.

No. Customer-voice agreement compares against saved model decisions rather than human labels, and tied or missing references are excluded, so treat it as fidelity to prior outputs rather than ground truth.

With enough VRAM, pip install -e '.[inference]' and use the DiffusionHarness class, calling load() with your Hugging Face authentication and predict() with requests and a step count; local unit tests need no credentials or GPU.

Actual JSON with the chosen values, such as an object with department set to billing and refund_requested true, with no letter-code translation, supporting strings, integers and booleans.

No. It is an independent experiment that does not reproduce Jev's architecture, training, calibrated probabilities or published workflow scores, and its restricted softmax scores are uncalibrated rather than reliable truth probabilities.