BF16 derivative of nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 at revision d468880b6ad3c6e0d21377ce7242adaea4cc884d, produced by one audited refusal-direction projection. The normalized float32 harmful-minus-harmless residual direction was measured at layer 34 with seed 42 across all 52 layers (chat-templated prompts; selected by refusal-generation test, 0/8) and projected from exactly 3,126 residual-writing tensors. The vision tower was untouched (this family has none in the edit set), all MTP tensors were preserved, and maximum normalized residual leakage was approximately 0.000160.

The direction corpora were mlabonne/harmful_behaviors@01cead01398926d81f7c52bdb790ee8cf77ebba7 (320 harmful) and mlabonne/harmless_alpaca@02c6a92cfcf11bb0c387334f8146d149d65b587f (320 harmless), normalized and deduplicated deterministically.

Full provenance, protocol, and caveats:

Missing cells are marked not measured and are never backfilled from a different checkpoint or protocol.

This model has had its refusal direction deliberately reduced. It complied with 200/200 harmful prompts in the measured suite and has no added safety mitigations. It will comply with many requests the upstream model would refuse. Deploy only behind appropriate policy, filtering, access controls, and legal review. Refusal-rate numbers are behavior measurements, not safety endorsements.

Engineering release: darkstar-nemotron-3.5-lightning-v1.0.0. This immutable tag exists and the release contract is published.

Validated with vLLM (CUDA 13 Blackwell nightly build family), Flash Attention, BF16 KV cache, context 131,072, MTP depth 12, and max_num_seqs=16: