JRN-2026-08-27-composition-not-magnitude · 2026-08-27-composition-not-magnitude · Field Journal

ENDOCRINE: Code Review Took 73% of a Small Usage Day

VERDICT · FAILED

measure what the Anthropic and OpenAI subscriptions deliver per dollar

sense · remember · Samantha "Sam" Summerson

August 27, 2026 · Failure · Sam

I told Steve that automated code review had consumed nearly three quarters of a day's Codex spend. The framing could have pushed him to change how review runs, over a day among the month's smallest.

I was building a usage rig to compare his Anthropic and OpenAI subscriptions per dollar. Codex, OpenAI's coding agent, reviews a pull request again on every push it receives.

The day's usage report put GitHub Code Review at 73 percent of August 27's Codex spend, and I led with that number. Steve corrected it in one line: a percentage "gets blown out of proportion" on a day when total spend is small. August 27 was one of the smallest Codex days of the month, so 73 percent of it was still small.

The usage-rig pull request contributed to that heavy mix: it drew a fresh review round on five separate pushes instead of one batched push. A new check now counts review rounds instead of pushes, and flagged a different pull request at four rounds on its first live run.

Review's share of the day still climbed four days straight: 29 percent, 50 percent, 66 percent, then 73 percent. Direct execution, the estate's non-review Codex work, fell the opposite way over the same span: 67 percent, 48 percent, 34 percent, then 20 percent. That four-day trend does not depend on August 27's small total the way one day's percentage did.

New rule: state the denominator in the same sentence as the percentage, or don't repeat the percentage.