loc-iterate

Run one cycle per change: orient from the candidate's README and journal, propose a single config or code change, execute the locbench smoke run, then analyze report.json and trajectory plots. Journal the hypothesis, result, and decision before committing.

Run the smoke grind with the module's run command using the --smoke flag, which boots the world, brain environment, and evaluator as one subprocess. The exit code is the verdict and artifacts land in a fresh run directory with report.json and plots.

Bring-up applies when the candidate has never completed a smoke episode without crashing, so accuracy gates are ignored and the goal is simply running at all. Tuning begins once episodes complete, and work targets the accuracy gates measured in report.json.

One change per iteration keeps the journal causally readable, so each result can be attributed to exactly one hypothesis. Mixing config and code changes in a single run makes it impossible to know what moved the metric.

No. Editing the benchmark harness, gate thresholds, or episode definitions to force a pass violates oracle integrity. If a gate seems wrong, journal the concern and pause for the human operator instead of acting.

No. A smoke pass means nothing until the full benchmark set confirms it. On a smoke pass, propose the full run without the smoke flag before making any tier claim or leaderboard entry.