Task Repo Guide

    Your task repo

    Ops grants you access per task; the link is on your HAI task page.

    Definitions of every term used here are on the Terminology page.

    Conventions

    • Only files under submissions/ belong in the PR, plus envs/tasks/<prompt_id>.dockerfile and .md if you changed them; hs submit includes those for you. Your own checks go in verification/armA/ and verification/armB/; detailed reasoning goes in grading evidence. No notes.txt or other loose files.
    • Your task repo is read-only. Fork it, clone your fork, and push with hs submit --push --pr (see Step 3 and Step 6).
    • Explain any steering deviation between the arms in the comparison comments, citing the turns.
    • A rerun keeps the earlier run under submissions/previous-runs/; give the reason when hs asks.

    Task types, domains, and interaction modes are on About the Task.