TouchStoneRL environment CI
Stack comparisons Connecting…

A reward is not a completed task.

Run a grader against verified controls and cheating attempts. Compare the reward with independently checked behavior before an environment enters training.

Start with an intentionally vulnerable example, then audit its repaired grader. You can also import your own Python coding-task package.

New environment audit

Select an environment package. The runner freezes it and resolves its image identity before probing. Edit the starter JSON to change cases, candidates or policy.

Task and audit boundary

Candidate probes

    Audits execute locally with bounded resources and no workload network. Candidate and grader containers are separate. The grader receives grade fixtures; candidate code never receives expected outputs or the grader. Answer graders have no candidate container: responses are generated strings.

    Import an environment

    Start from the repaired environment package. Supply your grader, reference implementation, separate grade/audit fixtures, and invalid or attack candidates. For a reward function that scores response strings, draft an answer-grader package with touchstone-rl init and import that.

    Maximum 1 MiB per package. Import registers code locally; it does not execute it. You review and start the audit separately.

    Probe observation

    Loading observation…