back when i was wiring up a toy prompt scrubber on a ragged little gpu box, the embeddings looked harmless enough in my notebook until i tried to reconstruct a few “sanitized” customer prompts and kept getting the same product names back. i remember one case where a field note with a weird typo still leaked cleanly enough that the nearest-neighbor dump was basically laughing at me. after that i stopped trusting anything that only blurred the representation a bit, because the thing that saved the downstream task was usually the same structure that made inversion annoyingly easy. obfuslm style schemes always felt a little too much like that to me.