v1.5.0

New abliterated models

• gpt-oss-120b (74% ASR) via vLLM TP in-place editing for MoE
• Qwen3.6-27B V1 (16/100) and V2 (10/100, manual iterative peel)
• Gemma-4-31B (3/100) via vLLM in-place direct-edit recipe

Engine

• min_gen_tokens plumbed through HF / vLLM / sglang / scorer / detector
• vLLM continuation-NLL KL path (score_continuations_nll) for in-place edits — fixes the KL=0.0000 sampler bug on Gemma 4
• Gemma4 model_type compat (vllm_compat.py + vllm_hidden_states whitelist with JSON fallback for unknown text_configs)
• Blackwell PCIe deadlock fix: disable_custom_all_reduce=True
• FP8 NaN/clamp guard in vllm_moe_editor after projection
• CLI / optimizer support for disable_lora=true and attention-only direct edit

Evaluation

• New scripts/eval_local_refusal.py: local Mac/MPS or OpenAI-compatible refusal evaluation
• scripts/eval_model.py: --eval-set {target,benign,both}
• Any OpenAI-compatible endpoint can be the LLM judge (#18)
• MiniMax added as an alternative judge provider (#17)

Recipes & docs

• New docs/vllm.md runbook
• configs/minimax_m2.7_vllm.toml LoRA + zero-companion expert path
• Gemma-4-31B 3/100 recipe shipped
• README / SPEC / configuration / evaluation / runpod-deployment docs updated

Housekeeping

• /artifacts/ and /results/ runtime output dirs added to .gitignore
• ProjectionCache regression test for plain Linear without LoRA base_layer

Full Changelog: v1.4.0...v1.5.0