Zoo Leaderboard Score View Implementation Plan¶

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

Goal: Add a higher-level leaderboard score-axis switch so users can choose return-based or normalized-score-based ranking without spelling raw metric aliases.

Architecture: Extend src/rl_training/zoo_cli.py with a leaderboard-only --score-view return|normalized option that resolves through the existing compare-to/metric alias layer. When no explicit --leaderboard-metric or raw --sort-by is provided, score_view should choose the return or normalized axis while compare_to still selects best vs latest. Carry the resolved score_view into leaderboard payload metadata and reject score_view=normalized when the manifest has no score normalization configured.

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

Task 1: Add failing regression tests¶

Files: - Modify: tests/test_cli.py

Step 1: Write the failing tests - Add a leaderboard JSON test verifying --compare-to latest --score-view return on a normalized benchmark resolves to latest-return and sorts by return instead of normalized score. - Add a leaderboard test verifying --score-view normalized on an unnormalized manifest raises a clear error.

Step 2: Run test to verify it fails

Run: pytest -q tests/test_cli.py

Expected: FAIL because --score-view is not supported yet.

Task 2: Implement score-view resolution¶

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

Step 1: Write minimal implementation - Add --score-view parser support. - Resolve score_view through the existing leaderboard metric alias layer only when no explicit --leaderboard-metric or raw --sort-by is provided. - Surface score_view in leaderboard payload metadata and text/CSV outputs. - Reject score_view=normalized when the manifest lacks score normalization.

Step 2: Run focused tests

Run: pytest -q tests/test_cli.py

Expected: PASS.

Task 3: Document score-view usage¶

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

Step 1: Add docs - Show --score-view return. - Explain that --score-view controls the return vs normalized axis while --compare-to controls best vs latest.

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.