Divergence Alarm
A scale-free statistic between the preference posterior and the reality posterior that rings when the two stop agreeing under continued optimization. Three estimators in increasing cost: Kendall-tau between ranks, regret of the preference winner under reality, and the slope-opposition trend where the proxy climbs while realized outcomes turn down.
Why It Exists
Goodhart's law is unfalsifiable when there is only one number; you cannot detect the proxy being gamed unless something else exists to compare against. The alarm is the only thing that fires on reward hacking before it shows up in the P&L - because by the time it shows up in the P&L, the optimization has already converged on the wrong target.
Related Terms
Drift Detector - A posterior predictive check that detects when the operator's preferences have drifted from the learned model.
Verification Trap - A task that is easy to generate but hard to verify.