Representational change under post-hoc edits mostly reflects relocation rather than destruction of information

Trained representations are edited by continual learning, model editing, and concept erasure, and a large change is commonly read as evidence that information was destroyed. We test this reading on testbeds with ground-truth factors (a β-VAE on dSprites; convolutional classifiers on Split-MNIST and CIFAR-10), separating information moved to new coordinates from information that is gone, using frozen versus fresh probes and latent-health metrics (effective rank, coordinate scale). Two linear edits, belief revision and LEACE erasure, share one dominant outcome: the original readout breaks, yet a fresh probe recovers the factor. Under belief revision at a healthy step size, permanent loss averages 1–2 percentage points over factors; the result transfers to real convolutional networks, although how much drift a cheap affine map reverses is substrate-dependent. Relocation is not universal: under linear erasure, nonlinear leak-back ranges from 0.54 to 1.00 over 32 cells, and scale loses up to 46% of its nonlinearly recoverable information at a healthy latent. In our experiments, total loss occurs only through representational collapse under an excessive optimizer step. Drift correlates only modestly with loss within a sweep and does not order outcomes across configurations; change magnitude is a poor proxy for loss.