Use an outcome your team accepts
Count a task when it clears the frozen acceptance policy, required checks, and review conditions, because generated tokens, changed lines, started sessions, and a green model response do not qualify on their own.
Token price has no quality denominator. Fullbeam reports observed attempt cost per accepted-task equivalent and lists steering, retries, review, and rework separately, including when those fields are unavailable.
Each result stays tied to the Stack Release, workload, repository state, and coverage that produced it. If evidence is missing, the gap remains visible in the release call.
Count a task when it clears the frozen acceptance policy, required checks, and review conditions, because generated tokens, changed lines, started sessions, and a green model response do not qualify on their own.
Maintenance can be cheap while migrations consume retries and reviewer time, so report the cost by task class before combining anything into a portfolio estimate.
Report intervention turns, corrective instructions, restarts, CI repair, review revisions, rework, fallback calls, and abandoned attempts only when they are joined to the comparison. Otherwise mark them unavailable and excluded.
When baseline spend, eligible volume, runtime attribution, or accepted-outcome coverage is thin, the saving remains unavailable until better evidence arrives.
A candidate has lower observed accepted-task cost when it repeatedly produces equivalent accepted work for less measured attempt cost. A total-engineering-cost claim waits for complete effort coverage.