analyze-ablation

Analyze Ablation Results

When to use

• User says "analyze the ablation results" or "analyze ablation for "
• User wants to know which guideline sections to keep or remove
• User asks about guideline compaction or AGENTS.md token budget
• After a guideline ablation experiment has been run (via GitHub Action or locally)

Step 1: Find available results

List the ablation/results/ directory to see which models have results:

ls ablation/results/

If the directory is empty or doesn't exist, the user needs to either:

1. Run the ablation locally: bun run scripts/runAblation.ts --model <model>
2. Or download results from a GitHub Actions run:

   # List recent ablation workflow runs
   gh run list --workflow=ablation_experiment.yml --limit=5
   # Download the artifact
   gh run download <run-id> -n ablation-<model>-<run-id> -D ablation/results/
   

Step 2: Read the summary JSON

For the requested model, read the latest JSON file from ablation/results/<model>/. The file contains:

• model: The model name
• timestamp: When the experiment ran
• baseline: Overall pass/fail counts and per-eval results with the full guideline set
• sections: Array of per-section ablation results, each with:
  • name: Section name (e.g. "function_guidelines", "query_guidelines")
  • tokensInSection: How many tokens this section costs
  • verdict: "ESSENTIAL" (2+ regressions), "MARGINAL" (1 regression), "DISPENSABLE" (0 regressions)
  • regressions: Eval names that flipped from pass to fail when this section was removed
  • improvements: Eval names that flipped from fail to pass when removed (guidelines confusing the model)
  • score: Pass/fail counts for this ablation variant

Step 3: Present the classification table

Present a summary table showing:

Sort by verdict: ESSENTIAL first, then MARGINAL, then DISPENSABLE.

Step 4: Detailed analysis

For each ESSENTIAL and MARGINAL section:

• List the specific evals that regressed
• Check if the regressed evals are related to the section topic (e.g. query_guidelines causes query eval regressions — expected)
• Flag any surprising regressions (section removal caused failures in unrelated evals)

For any section with improvements:

• These are cases where the guideline was actively confusing the model
• Flag these as candidates for rewording even if the section is kept

Step 5: Cross-model comparison (if multiple models have results)

If ablation/results/ has results for multiple models:

• Compare verdicts across models
• A section ESSENTIAL in ALL models is definitely essential
• A section ESSENTIAL in some models should still be kept (conservative approach)
• A section DISPENSABLE in ALL models is safe to remove
• Flag any sections with conflicting verdicts across models

Step 6: Token budget summary

Calculate:

• Current full guidelines: Total tokens (from baseline)
• Essential sections only: Sum of tokens for ESSENTIAL sections
• Essential + Marginal: Sum of tokens for ESSENTIAL and MARGINAL sections
• Savings: How many tokens saved by removing DISPENSABLE (and optionally MARGINAL) sections

Step 7: Recommend next steps

Based on the results, recommend one of:

1. Ready to build AGENTS.md: If the classification is clear and token savings are meaningful, suggest building the compact guideline set and running a validation run.

2. Subsection ablation needed: If function_guidelines is ESSENTIAL (likely — it's the largest section at ~2400 tokens), suggest a follow-up ablation of its 8 subsections to find further savings.

3. Cross-model validation needed: If only one model has been tested, suggest running ablation on 1-2 additional models for confidence.

4. Results are noisy: If many sections show exactly 1 regression (MARGINAL), the run-to-run variance may be too high. Suggest re-running or using a different model.

Present findings to the user and ask which direction they want to go. Do NOT make any code changes until asked.

Reference: Guideline sections

The 10 top-level sections in runner/models/guidelines.ts are:

1. function_guidelines — Function syntax, HTTP endpoints, validators, registration, calling conventions, function references, API design, pagination
2. validator_guidelines — v.bigint deprecation, v.record usage
3. schema_guidelines — Schema location, system fields, index naming, index field ordering
4. typescript_guidelines — Id types, Record types, strict typing, as const, Array/Record patterns, @types/node
5. full_text_search_guidelines — Search index query syntax
6. query_guidelines — No filter, no .delete(), .unique(), async iteration, ordering
7. mutation_guidelines — ctx.db.replace vs ctx.db.patch
8. action_guidelines — "use node", no ctx.db, action syntax
9. scheduling_guidelines — Cron syntax, FunctionReference usage, crons.ts patterns
10. file_storage_guidelines — Storage API, getUrl, system table queries, Blob handling