What testing strategies would you implement to ensure a newly written ML kernel produces mathematically correct outputs?
