Zoo Leaderboard Stability Metrics Implementation Plan¶

For Claude: REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.

Goal: Make benchmark comparisons more robust by exposing cross-seed stability statistics and adding leaderboard modes that rank by seed-to-seed consistency.

Architecture: Extend aggregate benchmark rows in src/rl_training/zoo_cli.py with latest-metric stability summaries across seeds, specifically min, max, and std for return and normalized score. Reuse those fields in the existing leaderboard metric alias layer by adding stability-return and stability-normalized modes that sort ascending on standard deviation so lower variance ranks higher while single-seed groups remain unrated (None) and fall to the end.

Tech Stack: Python 3.10+, argparse, existing zoo CLI/report serializers, pytest.

Task 1: Add failing regression tests¶

Files: - Modify: tests/test_zoo_presets.py - Modify: tests/test_cli.py

Step 1: Write the failing tests - Add a report JSON aggregate test verifying latest_*_min, latest_*_max, and latest_*_std fields are emitted. - Add a leaderboard JSON test verifying --leaderboard-metric stability-normalized sorts by normalized-score standard deviation ascending and reports the resolved metric metadata.

Step 2: Run test to verify it fails

Run: pytest -q tests/test_zoo_presets.py tests/test_cli.py

Expected: FAIL because the new aggregate fields and stability aliases do not exist yet.

Task 2: Implement stability aggregation and ranking¶

Files: - Modify: src/rl_training/zoo_cli.py - Modify: src/rl_training/cli.py

Step 1: Write minimal implementation - Add helpers for aggregate min and std on numeric fields. - Emit latest return / latest normalized min, max, and std aggregate fields. - Extend leaderboard metric aliases with stability-return and stability-normalized. - Make stability aliases sort ascending because lower standard deviation is better.

Step 2: Run focused tests

Run: pytest -q tests/test_zoo_presets.py tests/test_cli.py

Expected: PASS.

Task 3: Document stability modes¶

Files: - Modify: README.md - Modify: zoo/README.md - Modify: src/rl_training/assets/zoo/README.md

Step 1: Add docs - Show --leaderboard-metric stability-normalized. - Explain that stability modes rank lower cross-seed standard deviation higher.

Task 4: Verification¶

Run: - Focused: pytest -q tests/test_zoo_presets.py tests/test_cli.py - Broader: pytest -q

Notes: - This plan intentionally omits commits because the session instructions forbid committing unless explicitly requested.