nx_activation_steer.nx

nx_activation_steer.nx -- activation steering primitive. First patch-composted-into-substrate primitive per cardinal feedback-loras-and-negatives-are-patches-not-systems (patches-as-manure refinement 2026-05-16): User: "im not against patches but patches are manure that should become architectural decisions to increase capability speed etc in the native stack" THE patch we're composting: - Negative prompts ("ugly, deformed, blurry, bad anatomy") -- a shotgun list of anti-features that the model never had measurable knowledge of. Folk-knowledge attractors. THE substrate capability we ship instead: - Per-layer ACTIVATION STEERING (Turner 2023 ActAdd / Zou 2023 Representation Engineering). Given a "steering vector" -- the difference between activations on two contrasting prompts at a chosen layer -- ADD it to forward-pass activations to bias output in the desired direction. Architectural advantages over negative prompts: 1. Per-token control. Steering is applied at every forward pass, not just at sampling time. No "first token escapes negative prompt" failure mode. 2. Composable. Multiple steering vectors can be added (e.g. "less violence + more cheerful + more detailed") with per-vector strength weights. 3. Measurable. The substrate can VERIFY the steering had the intended effect by composing nx_monte_carlo + nx_multivariate over generations with vs without steering applied. 4. Patent-clean. Vector arithmetic; no LoRA training, no negative-prompt grimoires. ===== Math ====================================================== Given hidden activations h [n_tokens, hidden_dim] at some layer L,