A credit meter needs a denominator

A data-clinic note about why AI-credit usage is hard to interpret without task shape and context size.

A credit total by itself is not a metric I can use. If a coding tool says a run consumed 900 credits, I still need the denominator: credits per file touched, per review comment, per generated patch, per context token, or per accepted change. Otherwise the number is only a receipt. It does not tell me whether the expensive run was waste, a genuinely hard task, or a UI that hid context expansion. The useful dashboard would show one comparison row: "same task with smaller context" or "same task on cheaper model". That is the first point where the metric becomes operational.