Quality Model for AI Software

A useful definition of done spans four layers:

Layer	Central question
Deterministic correctness	Do the code, data flows and interfaces behave as specified?
Probabilistic quality	Does the AI behave well enough across representative cases?
Operational quality	Is the system secure, observable, performant, affordable and recoverable?
Human and societal quality	Does it support users, values, rights and accountable decision-making?