| name | experiment-plan |
| description | Turn a refined research proposal into a claim-driven experiment roadmap with progressive 4-stage execution order. Reads from research-refine output (FINAL_PROPOSAL.md). Produces EXPERIMENT_PLAN.md + EXPERIMENT_TRACKER.md. Use after research-refine, or when user says "实验计划", "experiment plan", "设计实验", "ablation matrix", "evaluation protocol", "run order". |
Experiment Plan: Claim → Evidence → Run Order
Turn method proposal into a claim-driven experiment roadmap. Every experiment must change a reviewer belief or it gets cut.
Domain-specific details (VLA sim-to-real protocols, Peng Sida 十问, etc.) are in references/domain-guide.md.
Constants
- MAX_CLAIMS = 2 — one dominant + one supporting
- MAX_BLOCKS = 5 — compact paper story
- MAX_BASELINES = 3 — few strong > many weak
- DEFAULT_SEEDS = 3
- OUTPUT =
refine-logs/ — shared with research-refine
Pipeline
refine-logs/FINAL_PROPOSAL.md
│
├─ P0: Load proposal (Problem Anchor, method, constraints)
├─ P1: Freeze claims & anti-claims
├─ P2: Build 5-block experimental storyline
├─ P3: Specify each block (dataset, baselines, metrics, success criteria)
├─ P4: 4-stage progressive run order
└─ Output: EXPERIMENT_PLAN.md + EXPERIMENT_TRACKER.md
P0: Load Proposal
Read refine-logs/FINAL_PROPOSAL.md. Extract: Problem Anchor, dominant contribution, method details, compute/data constraints, reviewer-flagged risks.
P1: Freeze Claims & Anti-Claims
| # | Claim | Anti-Claim to Rule Out | Minimum Convincing Evidence | Blocks |
|---|-------|----------------------|----------------------------|--------|
| C1 | [dominant] | "gain is just from more params" | method > baseline by ≥X on Y | B1,B2 |
| C2 | [supporting] | "component is decorative" | −component < full by ≥X | B3 |
Anti-claims are as important as claims. No claim without an anti-claim. Evidence must cite specific metric + threshold. Max 2 claims.
P2: Build 5-Block Storyline
| Block | Question | Type |
|---|
| B1: Main Anchor | Does the method solve the bottleneck? vs strongest baselines on primary benchmark | Main paper |
| B2: Novelty Isolation | Does the dominant contribution itself matter? Ablation removing only the novel component | Main paper |
| B3: Simplicity Check | Can a bigger version be avoided? Overbuilt variant comparison | Main or Appendix |
| B4: Frontier Necessity | Is the modern primitive actually needed? vs strongest simpler alternative | Main or Appendix |
| B5: Failure Analysis | What does the method still miss? Error taxonomy, qualitative diagnosis | Main paper |
Delete blocks that don't apply. Add domain-specific blocks if needed (see references/domain-guide.md).
P3: Specify Each Block
For each kept block:
### B<N>: [Name] — [MAIN/APPENDIX]
**Claim**: C1/C2 | **Anti-claim**: [what this rules out]
**Dataset**: [name, split, task formulation]
**Compared**: Strongest baseline | Simpler variant | Ours (full) | Ours (−component)
**Primary metric**: [must show clear improvement]
**Setup**: backbone, key hyperparams, GPU-h/run, seeds
**Success**: [specific threshold = "convincing"] | **Failure means**: [interpretation]
**Priority**: MUST-RUN / NICE-TO-HAVE
P4: 4-Stage Run Order
| Stage | Goal | Gate | Cost |
|---|
| M0: Sanity | Overfit tiny split, verify pipeline | Loss decreases, no NaN | < 0.5h |
| M1: Baseline | Reproduce strongest baselines fairly | Match published ±σ | Xh |
| M2: Main | Full method + novelty ablation (B1+B2) | Claim holds | Xh |
| M3: Decision | Simplicity + frontier checks (B3+B4) | No bloat | Xh |
| M4: Polish | Multi-seed, failure analysis (B5), appendix | All MUST-RUN done | Xh |
If Stage M2 fails: diagnose via Peng Sida 十问 (see references/domain-guide.md).
Output
EXPERIMENT_PLAN.md
# Experiment Plan
**Problem**: [...] | **Method Thesis**: [...] | **Date**: [...]
## Claim Map (table)
## Paper Storyline (B1-B5, main/appendix/cut)
## Experiment Blocks (B1-B5 full specs)
## Run Order (M0-M4 table with gates, GPU-h, risks)
## Compute Budget (total GPU-h, data prep, bottleneck)
## Risks & Mitigations
EXPERIMENT_TRACKER.md
| Run ID | Stage | Purpose | System | Dataset | Metric | Target | Priority | Status |
|--------|-------|---------|--------|---------|--------|--------|----------|--------|
| R001 | M0 | sanity | Ours | toy-5 | loss↓ | <0.01 | MUST | TODO |
Key Rules
- Every experiment defends or rules out a claim. No claim → cut.
- Compact story. Main table first, add ablations only as needed.
- Defend simplicity. Include deletion study.
- Strong baselines > long lists. 3 well-tuned > 10 weak.
- MUST-RUN vs NICE-TO-HAVE. Appendix must not delay core evidence.
- Stage M0 sanity saves weeks. Overfit first, scale later.
- Never fabricate results. Plan evidence; don't claim evidence.
Calls
- Reads: refine-logs/FINAL_PROPOSAL.md (from research-refine)
Called by
research-refine → experiment-plan (YOU ARE HERE) → paper-write