Directional semantic coherence score (DSCS)

Fits a linear direction separating two semantically opposite action classes (e.g. push vs. pull) from frozen video-model activations, then measures the consistency of that direction's sign/orientation across held-out clips, distinguishing an oriented axis of variation from mere cluster separability (which cosine-similarity or accuracy alone can conflate).