Definition
The full cost of producing, validating, correcting, and operating work divided by outcomes that pass acceptance and remain useful for a defined period. Dark Factory Dev uses it as a practical unit metric for comparing agent systems whose cheap first attempts can hide expensive review, retry, incident, and maintenance work.
For a stated measurement window:
cost per accepted durable outcome = total attributable cost / number of accepted outcomes still durable at the checkpoint
The numerator can include model usage, compute, tools, sandbox infrastructure, evaluation, retries, human review, rework, incident response, and the opportunity cost of scarce human attention. The denominator includes only outcomes that passed a defined acceptance process and still met the durability rule later.
Define it before measuring it
The metric is only comparable when the unit of outcome, acceptance gate, durability period, risk class, and cost boundary are explicit. A typo fix and a payment migration are not equivalent outcomes. A seven-day window and a one-year window do not measure the same durability. Teams should report those choices beside the number rather than presenting a universal dollar figure.
Distinguish it from nearby terms
Token cost measures one resource. Task-success rate measures immediate completion against some test. Cost per accepted durable outcome combines cost, acceptance, and survival over time. It is related to outcome-maxing, but it is an accounting measure rather than an optimization slogan.
Check your understanding
System A spends $1 per attempt and produces 100 attempts, but only 50 pass review and 40 remain valid after 30 days. System B spends $2 per attempt on 60 attempts, with 55 accepted and 54 still valid. Which is cheaper? You cannot answer from per-attempt price. Apply the same cost boundary and durability rule to both systems.