Reveneau Claude Code token benchmark, first dated run: 4 October 2026 (Pacific time; the last runs finished after midnight on 5 October).
runs.json / runs.csv: every run, one row each, as reported by `claude -p --output-format json` (total_cost_usd, input, output, cache_read, cache_write, num_turns), plus pass/fail from re-running the test after the run, and wall-clock seconds.
task/: the repository the runs worked in (three source files, one test file with one failing test), the CLAUDE.md used in the rules experiment (301 lines of text, 342 counting blank lines, 62,746 characters), and the noisy test used in the hook experiment.
scripts/: run.sh (rules experiment), run2.sh (model, effort and hook experiments), the hook script and its settings file.
Claude Code 2.1.118 on macOS, signed in with a Claude subscription. Model aliases resolved to claude-sonnet-4-6, claude-opus-4-7 and claude-haiku-4-5-20251001 on that day.
One earlier run was discarded before the five recorded no-rules runs: its test command named a directory, which Node rejected, so pass/fail could not be read. It is not in the data.