AI governance
Systems that know when they do not know
A softmax classifier cannot output 'unknown'. It is obliged to guess, and it will assign high confidence to inputs it has never seen. In embedded infrastructure that is not a statistical curiosity, it is a liability.
A recommendation engine that misclassifies a film causes mild annoyance. A vision system that misclassifies a signal, a gauge, or a person does not. The difference is not the model architecture. It is the consequence attached to a confident wrong answer.
Standard classifiers return a distribution over the classes they know. Presented with something outside that set they do not return nothing; they return the nearest thing, with a confidence score indistinguishable from a correct answer.
Engineering the reject option
The fix is to design abstention in deliberately: a threshold above which the system declines to answer and escalates instead. Conformal prediction is one practical route, because it gives a statistically grounded bound rather than an arbitrary cut-off on a softmax score.
That requires separating uncertainty arising from noise in the input from uncertainty arising because the model is outside its training distribution. The two call for different responses — one is tolerable, the other means the system is operating somewhere it was never validated.
The operational value of not answering
Abstention has a measurable return: the cost of the wrong actions the system did not take, less the cost of the human time spent on the cases it referred. On any deployment where a false positive triggers a physical intervention, that arithmetic is rarely close.
This is the L2 Readiness question from the framework we publish, asked from the other end. Readiness asks whether the data can carry the decision. Abstention asks what the system should do on the days it cannot.