$ ls data/

data/

numbers from our own runs, derived in the open, dated and sourced

as of Sep 7, 2026jsonClaude Code agent run costs: recorded tokens and modeled API-equivalent pricingRecorded token fields from devweb agent-workflow runs, with dated API-equivalent pricing scenarios and coverage limits.17 min

$ ls evidence/

limited evidenceDoes an AI code reviewer catch real problems?The review stages logged 135 coded findings across 8 runs, 31 of them BLOCKER and 92 coded FIXED, plus 25 cold-review escalations and 5 proofreader holds; nothing here shows what share of real problems that was.8 minmixed evidenceHow often does a critic loop change the output?A critic loop fired at least once in 49 of the 67 runs with a record, a 73% share, mostly at the architect and prompts stages, while only 23 of 61 runs recorded any rework spend at all and the median was 0.0.8 min