An uncensored 27B cyber LLM. Qwen3.5-27B base — abliterated, then LoRA-trained on offensive trajectories. No filter, no leash: it answers anything.
Arditi et al: refusal lives in a one-dimensional direction of the residual stream. B.R.O removes it from every projection — no RLHF re-run, no runtime filter, nothing to bypass at inference time. Then offensive trajectories are distilled in on top, LoRA rank-16, and the adapters fuse back into the base: one model, one file set.