A change in direction.
A result with limits.

What this result establishes

Adding a fixed hidden-activation vector causes a measured change in world-Y motion for the evaluated checkpoint. A change along the world’s Y axis does not by itself establish body-relative sidestepping, a clean semantic concept, or a reliable locomotion controller.

Projection ablations and the saved observation transformation support a heading-related explanation. That transformation rotates already-clipped contact wrenches too: exact physical-heading equivalence and unique heading mediation remain unestablished.

From observations to a causal test

1. Collect

   Unsteered rollouts from a frozen pretrained policy.

2. Extract

   Episode-balanced differences between high and low velocity activations.

3. Lock

   Select vector strength using validation episodes.

4. Compare

   Paired held-out episodes, controls, failures and later mechanism audits.