How do you deal with Correction Cascades if you need to have certain hard constraints for the ML?
Chris Butler

As with regular software anti-patterns it is not fully prohibited but there is a price to pay. In case of ML anti-patterns the price could be very high but you weight it against the price of your ML system producing inferences that go against the regulations. Unfortunately, I am not familiar with a general solution to correction cascades — it may require a case-by-case approach. Something you might try is: measure the negative impact of the correction cascade and if it is minuscule then let it be but keep it monitored; define the output domain of the ML within the constraint; heavily weight unwanted corner cases in the training data so that ML almost never infers them but when it does and gets corrected by the cascade it so rare that there is a small overall impact.