Goodhart's law

Every metric stands in for something you cannot measure directly. Handle time stands in for efficient service. Resolution rate stands in for problems solved. The correlation between the proxy and the real thing holds while nobody is pushing on it.

People push on it with restraint, because they know what the number is for. A support team told to reduce handle time will find efficiencies and mostly will not hang up on customers. A system optimising the same target has no such model. If shorter responses score better, responses get shorter, including the ones that needed length.

The gap between the proxy and the real thing is where the shortcut lives, and it is usually where the guardrail belongs. For handle time, the gap is rushing, so the guardrail is a quality or repeat-contact measure.

Gaming shows up in the shape of a distribution before it shows up in the mean. A system closing easy cases fast and leaving hard ones has a stable average and a distribution pulling into two humps. Outputs that grow more similar to each other over time while the score holds is the same signal.