Decide which of your numbers is the control before you run the arm

scope: generic · severity: law · confidence: proven · subsystem: method

Every measurement needs a second number whose job is to prove the first one means anything.

The worked example. An IPA idle-observation arm watched power/runtime_suspended_time climb. It looked clean and would have been written up as a result. But power/runtime_active_time never moved, which meant the callback under test had not run once. susp_ms was the reading; act_ms was the control, and only the control was informative.

Decide which of your numbers is act_ms before you run the arm. Afterwards you will rationalise whichever number you happen to have.

Cheap controls, in rough order of how often they save you:

a counter that must increase if the code ran at all
the value read back from the device, not the value you wrote
a deliberately-broken arm that must fail (if it passes, your test is inert)
the total your filter matched before you count matches within it — see a-journal-grep-matches-your-own-command-line
ls /sys/module/<mod>/parameters/ before trusting anything a module parameter was supposed to change