Skip to content

Codex evaluation fixes, remove local file access, update results#25

Merged
justaddcoffee merged 4 commits into
mainfrom
codex-evaluation-dec2024
Dec 17, 2025
Merged

Codex evaluation fixes, remove local file access, update results#25
justaddcoffee merged 4 commits into
mainfrom
codex-evaluation-dec2024

Conversation

@justaddcoffee

Copy link
Copy Markdown
Contributor

Summary

  • Fix Codex cheating issue: Codex was accessing filesystem to read answer files. Now runs with disable_shell_tool: true to enforce MCP-only mode
  • Rerun Experiment 1: New Codex results without filesystem access
  • Rerun Experiment 2: Cross-model comparison with gpt-5.1-codex-max, gpt-5.1-codex, gpt-5.1-codex-mini
  • Update notebooks and figures: All analysis notebooks updated to use new result files
  • Fix CI linting: Configure ruff to ignore F821 in notebooks (cell execution order issue)

Test plan

  • Ruff linting passes
  • Notebooks execute without errors
  • Figures regenerated with new data

justaddcoffee and others added 4 commits December 15, 2025 10:51
- Update experiment notebooks with latest analysis
- Update pyproject.toml and uv.lock dependencies
- Remove obsolete figure files (exp1_*, exp2_*, failure_categories_*)
- Update .gitignore to exclude results/compare_models/ and scripts/

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Major changes:
- Update metacoder dependency to codex-disable-shell-tool branch
- Add disable_shell_tool: true to all Codex eval configs
- Rerun Experiment 1 (Codex) with MCP-only mode (38% pass rate)
- Rerun Experiment 2 (cross-model) with MCP-only mode:
  - gpt-5.1-codex-max: 54%
  - gpt-5.1-codex: 52%
  - gpt-5.1-codex-mini: 50%
- Update notebooks to use new result files
- Regenerate all figures with corrected data
- Remove old result files that had filesystem access (cheating)

Key finding: Model size has minimal effect (4pp spread) when
filesystem access is disabled, vs 11pp spread before.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Ruff does not understand Jupyter notebook cell execution order,
so it reports undefined name errors for variables that are defined
in earlier cells. This is a false positive for notebooks.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add missing outputs field to code cells (notebook schema fix)
- Remove unused f-string placeholders (ruff F541)
- Other auto-fixes from ruff --fix

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor
PR Preview Action v1.6.2

🚀 View preview at
https://contextualizer-ai.github.io/mcp_literature_eval/pr-preview/pr-25/

Built to branch gh-pages at 2025-12-17 20:43 UTC.
Preview will be ready when the GitHub Pages deployment is complete.

@justaddcoffee justaddcoffee changed the title Codex evaluation fixes: MCP-only mode and updated results Codex evaluation fixes, remove local file access, update results Dec 17, 2025
@justaddcoffee justaddcoffee merged commit 2f95260 into main Dec 17, 2025
5 checks passed
@justaddcoffee justaddcoffee deleted the codex-evaluation-dec2024 branch December 17, 2025 20:46
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant