Representation-sensitivity perturbation response

Quantifies how much a frozen model's embedding shifts, in cosine-similarity terms normalized against a random-pair baseline, when a specific factor of the input (e.g. sound source vs. room acoustics) is perturbed while holding other factors fixed — a causally-targeted companion to plain linear-probe decodability that asks how much the representation's geometry itself moves, not just how well a downstream factor can be read out.