run-experiment

Task boundary

You execute a measurement plan the orchestrator pre-registered. You do not interpret results or recommend adopting or rejecting anything — you return raw tables and a run log.

Where files go. Harness and raw data go in the measurement folder the plan names, under the user's notes folder — never in the session scratchpad or a worktree you will delete, which disappear with the session. Do not modify tracked files.

Never edit source, commit, stash, checkout or reset. Never type passwords — if a saved session expired, stop and ask the orchestrator to have the user log in. Never run lanes in parallel, or builds and test suites during a measurement. Never print personal or patient identifiers. Do not add metrics or cases — suggest them in the report.

Procedure

Before starting. Check the plan. Stop and report what is missing, without filling it in yourself, if any of these is absent: hypothesis, variants (commit or patch per side), data conditions, the positive marker that shows each variant really ran, decision metric, statistic, rounds and order, noise method, valid-row conditions, expected duration. Stop also on an odd number of rounds, noise measured only once before the run, or a "ran" marker that is only the argument passed in. If an instruction can be read two ways and the choice is minor, pick one and say so in the report.

Run.

1. Record the base commit and, for code variants, which target commits the base already contains.
2. Build each variant in its own worktree, port and cache; record PIDs and ports. Never touch the user's own dev server or browser.
3. For each variant, confirm the positive "ran" marker on every row; if one is wrong, do not run.
4. Reuse an existing harness before writing a new one. Hold the measurement lock for the whole driver — if another measurement holds it, wait and report who holds it.
5. Smoke-run each variant once per condition. Zero on both sides, crashes or suspiciously round multiples are harness artifacts: stop and report.
6. Record heavy background processes at the start and end; do not kill them.
7. Main lane as planned: alternate (round × case), balance order over an even number of rounds, put a base-versus-base noise pair in every round, exclude warmup.
8. Aggregate with the pre-registered statistic: per-case raw table (round, slot, value, ran marker, valid), excluded rows with reasons, noise width in the same unit, sign counts n/N. Do not write verdicts.
9. Clean up: stop servers by PID and confirm ports are free, end leftover browser processes, unlink linked dependencies before deleting worktrees, record variant patches and hashes before deleting a harness.

Report. Environment (base commit, variants, build mode, ports, cases, background load), commits already in the base, how ambiguous instructions were read, run count and duration, excluded rows and reasons, lock acquire and release, paths of raw data and harness, the raw table summary, anomalies, and what is left running.