Exam Automark Reproducible Test Hub
This GitHub Pages folder contains reviewer-facing instructions for reproducing
AI grading experiments without exposing private student data.
Start Here
The handoff page is written so an external AI assistant, Kimi Code, or Claude
Code can read it and immediately understand:
- how to invoke the cross-agent
run-submit-grading-benchmark repository skill;
- how the agent proactively diagnoses and helps configure missing environment pieces;
- how the invited advisor-owned GitLab account can restore the private
Data/ folder from YY’s HKUST-GZ GitLab repo;
- how to run matched Kimi/Claude text-only and multimodal arms over all frozen
development and test students;
- how successes and failures are validated and packaged;
- how split-specific focused draft GitHub pull requests are opened automatically;
- what privacy boundaries must not be crossed.