Ablation Planner

Systematically design ablation studies that answer the questions reviewers will ask. The reviewer agent leads the design; the local executor reviews feasibility and implements.

Context: $ARGUMENTS

Constants

• REVIEWER_MODEL = gpt-5.4 - Used via a secondary Codex agent for reviewer-style ablation design.

When to Use

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

Workflow

Step 1: Prepare Context

Read available project files to build the full picture:

• Method description and components (from docs/research_contract.md, project notes, or method docs)
• 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 server notes, run configs, or user-provided budget)

Step 2: Secondary Codex Designs Ablations

spawn_agent:
  model: REVIEWER_MODEL
  reasoning_effort: xhigh
  message: |
    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 or replaceable components]
    Current results: [key metrics from experiments]
    Claims: [what we claim and current evidence]

    For each ablation, specify:
    - name: what to change (for example, "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 will not add insight
    - suggested_order: run order optimized for maximum early information
    - estimated_compute: total GPU-hours estimate