Can a reward function tell a good lap from a bad one? An offline test, no simulator, no AWS. Found two real defects in 12 production DeepRacer reward functions.

An agent's claim that it finished is worthless unless the proof could have failed. A one-page method, no framework, no install.