Reproduce the scores

This preview provides aggregate verification and an offline JPT cache scorer. It does not rerun JPT inference. Public code access is pending; the commands below require the privately reviewed workspace package.

1. Install the reviewed anvil-hf-workspace package in a fresh Python 3.12 environment.
2. Obtain the author's archive and evaluator from JPT's own release. Compare exact SHA256 pins in the downloadable manifest and the package's reproduction source lock. No author artifacts are redistributed here.
3. Set ANVIL_HF_REPRO_RAW to your private author-file directory and ANVIL_HF_REPRO_OUTPUT to a new report directory, then run hf-workspace-reproduce. See the workspace's benchmarks/crossner/reproduce.md for exact commands and member checks.
4. Validate the prepared aggregate directory with hf-workspace-validate. This checks arithmetic and integrity; it does not establish inference accuracy.

Saved Graphiti and GLiNER predictions remain private; their inference cannot currently be independently reproduced from this preview. Hosted immutable Graphiti model identity and historical JPT checkpoint identity are unknown.

Final approved public installation/download links, external clean-user reproduction and Hugging Face deployment remain release gates.