The cost of a feature, measured
The first measurement is pending. The ruler is frozen and the baselines are committed; the numbers land here after the maintainer runs the agent ruler (make bench-agent-measure).
Methodology
The prompts of every scenario are frozen before the first measurement, one per side, with their SHA-256 committed in CHECKSUMS.txt; changing a prompt reopens the series.
The comparison is against a Go-pure app committed in this repository — the same feature, the standard library alone — so the denominator of every saving is something you can read.
Every number here is measured: the medians of the runs that came out green, three runs per side per scenario, tokens in (fresh plus cache reads) and out, from the agent's own result JSON. Nothing is estimated; when something is, it says est.
The saving is never stored — it is derived from the two sides wherever it is read, including here. The series file records what was spent, nothing else.
A release fails its gate when any scenario regresses more than five points against its previous measurement, when any scenario is under 60%, or when the average misses the milestone target.