Explain the number

A number is evidence only once you know what produced it.

• Before the run, name the result that would change a decision and the noise floor that result must clear.
• Explain a number by what bounds it: the resource or code path that limits it, found from a profile or counters taken during the run and mapped to source, not by reading code.
• List what else the number could be measuring (errors, skipped or cached work, code that never ran, an untuned side of a comparison, noise, a share too small to matter) and rule each out with evidence.
• A reported number carries its run count, its spread and its limiter.
• Trust an eval result only after checking that every run actually did the task, that the gap holds across trials and models, and that the scenario matters.