Goodhart's law

In ML work it appears as models, teams or reward models improving the reported number by cheap routes: biased low forecasts under MAPE, flagging nothing to keep precision high, tuning against the validation set. Defences are guardrail metrics that expose gaming, checks of the metric against cost, and rotating or auditing the target. Manheim and Garrabrant (2018) categorize several variants.