ablation-planner

Files (1)

• SKILL.md

SKILL.md (first lines)

# Ablation Planner

Systematically design ablation studies that answer the questions reviewers will ask. Codex leads the design (reviewer perspective), CC reviews feasibility and implements.

## Context: $ARGUMENTS

## When to Use

- Main results pass `/result-to-claim` with claim_supported = yes or partial
- User explicitly requests ablation planning
- `/auto-review-loop` reviewer identifies missing ablations

## Workflow

### Step 1: Prepare Context

CC reads available project files to build the full picture:
- Method description and components (from `idea-stage/docs/research_contract.md`, legacy `docs/research_contract.md`, or project CLAUDE.md)
- Current experiment results (from EXPERIMENT_LOG.md, EXPERIMENT_TRACKER.md, or W&B)
- Confirmed and intended claims (from result-to-claim output or project notes)
- Available compute resources (from CLAUDE.md server config, if present)

### Step 2: Codex Designs Ablations

```
mcp__codex__codex:
  model: gpt-6-astra
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    You are a rigorous ML reviewer planning ablation studies.
    Given this method and results, design ablations that:

    1. Isolate the contribution of each novel component
    2. Answer questions reviewers will definitely ask
    3. Test sensitivity to key hyperparameters
    4. Compare against natural alternative design choices

    Method: [description from project files]
    Components: [list of removable/replaceable components]
    Current results: [key metrics from experiments]
    Claims: [what we claim and current evidence]

    For each ablation, specify:
    - name: what to change (e.g., "remove module X", "replace Y with Z")
    - what_it_tests: the specific question this answers
    - expected_if_component_matters: what we predict if the component is important
    - priority: 1 (must-run) to 5 (nice-to-have)

    Also provide:
    - coverage_assessment: what reviewer questions these ablations answer
    - unnecessary_ablations: experiments that seem useful but won't add insight
    - suggested_order: run order optimized for maximum early information
    - estimated_compute: total GPU-hours estimate
```

### Step 3: Parse Ablation Plan

Normalize Codex response into structured format:

```markdown
## Ablation Plan

### Component Ablations (highest priority)
| # | Name | What It Tests | Expected If Matters | Priority |
|---|------|---------------|---------------------|----------|
| 1 | remove module X | contribution of X | performance drops on metric Y | 1 |
| 2 | replace X with simpler Z | value of learned vs fixed | drops, especially on dataset A | 2 |

### Hyperparameter Sensitivity
| # | Parameter | Values to Test | What It Tests | Priority |
|---|-----------|---------------|---------------|----------|
| 3 | lambda | [0.01, 0.1, 1.0] | sensitivity to regularization | 3 |

### Design Choice Comparisons
| # | Name | What It Tests | Priority |
|---|------|---------------|----------|
| 4 | joint vs separate matching | whether joint adds value | 4 |

### Coverage Assessment
[What reviewer questions these ablations answer]

### Unnecessary Ablations
[Experiments that seem useful but won't add insight — skip these]

### Run Order
[Optimized for maximum early information]

### Estimated Compute
[Total GPU-hours]
```

### Step 4: CC Reviews Feasibility

Before running anything, CC checks:
- Compute budget: can we afford all ablations with available GPUs?
- Code changes: which ablations need code modifications vs config-only changes?
- Dependencies: which ablations can run in parallel?
- Cuts: if budget is tight, propose removing lower-priority ablations and ask Codex to confirm

### Step 5: Implement and Run

1. Create configs/scripts for each ablation (config-only changes first)
2. Smoke test each ablation before full run
3. Run in suggested order, using descriptive names (e.g., `ablation-no-module-X`)
4. Track results in EXPERIMENT_LOG.md
5. After all ablations complete → update findings.md with insights

## Rules

- **Codex leads the design. CC does not pre-filter or bias the ablation list** before Codex sees it. Codex thinks like a reviewer; CC thinks like an engineer.
- Every ablation must have a clear `what_it_tests` and `expected_if_component_matters`. No "just try it" experiments.
- Config-only ablations take priority over those needing code changes (faster, less error-prone).
- If total compute exceeds budget, CC proposes cuts and asks Codex to re-prioritize — don't silently drop ablations.
- Component ablations (remove/replace) take priority over hyperparameter sweeps.
- Do not generate ablations for components identical to the baseline (no-op ablations).
- Record all ablation results in EXPERIMENT_LOG.md, including negative results (component removal had no effect = important finding).

From wanshuiyin/Auto-claude-code-research-in-sleep, MIT.

More from wanshuiyin/Auto-claude-code-research-in-sleep