Add production-observed acceptance alongside the controlled benchmark — 82.44% / 3.473 accepted tokens/step over 20h of agent traffic, with temperature and harness caveats. Additive; the benchmark table is unchanged and correct.

Add PROVENANCE.md - tensor-level provenance

Add NVFP4 weights, config, tokenizer, and trust_remote_code modeling files

Add model card, NOTICE, and both governing licenses

initial commit