aris-experiment-plan

Provide your refined proposal or method idea and the Skill freezes the paper claims, designs experiment blocks with datasets, baselines, and metrics, then outputs a run order with milestones. Results are written to refine-logs/EXPERIMENT_PLAN.md and a tracker file.

A good ablation matrix isolates the dominant contribution, defends simplicity against bloated variants, and justifies any frontier component like an LLM or diffusion module against the strongest simpler alternative. This Skill limits plans to a few strong baseline families rather than long padded lists.

Yes, it explicitly supports proposals with LLM, VLM, Diffusion, or RL-based contributions by including a frontier necessity check block. If the proposal is intentionally non-frontier, that block is skipped instead of forced.

Use research refinement first to stabilize the method and claims, then use experiment planning once the next question is what exact experiments to run and in what order. A combined pipeline skill exists if you want both in one request.

The plan reuses constraints from your proposal and does not invent budgets or fabricate results; it plans evidence rather than claiming it. You still need to execute the runs and interpret outcomes, ideally with a run-execution and review-loop skill.