run-discriminating-experiment-first
Run the discriminating experiment first
Two things are almost the same. One behaves, the other does not. The instinct is to fix by the strongest hypothesis, verify, and if that fails move to the next hypothesis. That loop has a terrible payoff curve: a guess costs one full verification round and returns information only when it happens to be right.
A discriminating experiment costs the same round and returns information either way, because you decide before running it what each outcome rules out. Swap the two things' positions. Trade one attribute between them. Run both variants side by side in one build. Whatever comes back, half the candidate causes are gone.
The rule of thumb: count the variables that still differ between the working and the broken side. If that count is greater than one, you are not allowed to write a fix yet.