exp-designlisted

# /exp-design > Given an idea (or a free-text hypothesis), design a complete experiment plan. > Claims are the core: scope the claims to validate across three dimensions — Target, Decomposition, and Threats. > Design four types of experiment blocks: baseline (reproduce baseline), validation (core verification), ablation (factor isolation), and robustness (stress testing). > Experiments are ordered by dependency with decision gates between stages (sanity fail → early stop). > Optional Review LLM review checks experiment plan completeness. All experiments are written to wiki/experiments/ with graph edges. ## Inputs - `idea`: one of: - A slug from wiki/ideas/ (e.g. `sparse-lora-for-edge-devices`) - A free-text hypothesis description (provide the experiment goal directly) - `--review` (optional): enable Review LLM review to check experiment plan completeness - `--budget <gpu-hours>` (optional): total compute budget cap (GPU hours), affects robustness experiment scope ## Outputs - `wiki/experiments/{slug}.md` — one page per experiment block (status: planned) - `wiki/graph/edges.jsonl` — new tested_by edges: experiment → claim - `wiki/ideas/{slug}.md` — updated linked_experiments field - `wiki/graph/context_brief.md` — rebuilt - `wiki/graph/open_questions.md` — rebuilt - `wiki/log.md` — appended log entry - **EXPERIMENT_PLAN_REPORT** (printed to terminal) — experiment block summary, run order, compute budget ## Wiki Interaction ### Reads - `wiki/ideas/{slug}.md` — idea's