Why intelligence systems fail outside the lab
Accuracy is only one property of a useful system. Operations decide the rest.
A model can be accurate and a system can still fail. Data arrives late. Interfaces create ambiguity. Incentives distort inputs. Operators invent workarounds.
Real intelligence systems are socio-technical. Their performance depends on how evidence moves, how decisions are made and how exceptions are handled.
The lab asks whether the model works. The field asks whether the institution can use it.
GF / END OF NOTE