openJev-verdict-2.0

The 77.10% top-1, 0.0636 Brier and 1.44% ECE figures are project-local, measured with the project's own evaluate script against a test receipt on a 2,000-decision held-out split of a public typed-decisions dataset. The 72.70% attributed to Jev 1.13.0 is a vendor baseline cataloged in Laya's published evaluation suite, not a rerun here, and Laya's own figures come from that same suite.

No, the repository holds two. Verdict is the 151M ModernBERT and GLiClass model evaluated on 231 public JevBench tasks, published separately as rlcd-modernbert-151m. Verdict 2.0 is the separate 149.6M architecture for typed software workflows, and its weights are tracked as Git LFS pointers in the repo's artifacts directory.

Inference only, with byte-identical weights. Three fixes: calibrators now auto-load and the five-option scope restriction is gone, candidates are templated as NLI-style hypothesis sentences, and the context budget drops from 1,024 to 512 tokens. On the measured tasks, hard-tier accuracy stayed at 36.9% while hard-tier ECE fell from 0.298 to 0.118.

At that size it runs in Chrome or Edge through WebGPU in under 600 MB, with decisions reported at roughly 20 to 25 ms. Licensing is messy: the README states Apache 2.0, but GitHub reports the repository as NOASSERTION, so confirm before shipping. Note also that a plain TF-IDF plus logistic regression baseline scores 66.10% at about 8 ms on that same suite.

It is a 149.6M-parameter ModernBERT-based encoder that evaluates Choice, Score, and Noul schemas in one non-autoregressive pass rather than generating text. Decisions take roughly 20 to 25 ms and return calibrated confidence, and the calibration is dual-channel: 1.44% ECE on the project's own split, with hard-tier ECE dropping from 0.298 to 0.118 after the inference fixes.

Verdict 2.0 weights are tracked as Git LFS pointers in the repository's artifacts directory, so a plain clone gives you pointers rather than the model, and the Verdict checkpoint is published separately on Hugging Face as heman10x/rlcd-modernbert-151m. The tokenizer loads from the usual Transformers from_pretrained path, so you need Git LFS installed for the 2.0 weights and an LFS-capable fetch for the Hugging Face one.

Yes, and it is the most useful number in the report: a plain TF-IDF plus logistic regression baseline scores 66.10% on the same suite in about 8 ms. That is the floor worth knowing before treating the 77.10% as impressive, and it is also a reminder that a decision task may not need a transformer at all.