Experiment Setup

Setup rules

• If the evaluation command is unclear, do not guess; surface the ambiguity.
• If the long job has no meaningful preflight command, record that limitation explicitly and treat the future run as higher-risk exploratory work unless the user approves otherwise.
• If the preflight command or pass criteria are unclear, do not treat the run as launch-ready.
• If the metric parser is unclear, do not guess; no launch is ready until the parser rule is explicit.
• If the baseline is unknown, record it as TBD and propose a baseline-confirmation step.
• If expected outputs or validation checks are unclear, do not treat the run as launch-ready.
• If the user gives multiple ideas, separate:
  • approved now
  • later options
  • explicitly out of scope
• Keep the setup inspectable. Never convert it into hidden autonomy.
• If the task is blocked on missing information, say exactly what is missing.
• Do not run anything yet.

Output

Produce:

• a draft or saved research/project_brief.md
• a draft or saved research/experiment_plan.md
• a proposed baseline E000 objective
• the preflight status needed before a long run can become launch-ready
• a short list of what still needs approval

References and templates

• Use assets/project-brief-template.md.
• Use assets/experiment-plan-template.md.
• Use assets/todo-template.md when initializing research/TODO.md.
• Use assets/experiment-index-template.md when initializing research/experiment_index.md.
• Use references/setup-checklist.md when deciding what is still missing.
• For the structured run contract, read ../experiment-launch/references/runtime-contract.md.