[snimu]

Adds the [tool.verifiers.eval] convention in preparation for the automated PR-eval action:

• AGENTS.md: every environment declares its full-eval convention (num_examples, rollouts_per_example) in pyproject.toml; automated eval runs read it when no sample count is given; infinite tasksets fall back to 50 examples.

Same convention as prime-envs already uses (e.g. environments/multimodal/mmk12).

---

Note

Low Risk
Documentation-only change to AGENTS.md with no runtime or config edits in this PR.

Overview
Documents the [tool.verifiers.eval] convention in AGENTS.md so contributors know how full evals are configured before automated PR-eval runs land.

The new rule states that every environment must set num_examples and rollouts_per_example in pyproject.toml. When automated eval runs omit an explicit sample count, they use those values; infinite tasksets default to 50 examples.

^{Reviewed by Cursor Bugbot for commit 6476f69. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Document the [tool.verifiers.eval] convention for environment eval configuration

Updates AGENTS.md to require every environment to declare its full-eval settings in pyproject.toml under [tool.verifiers.eval], specifying num_examples and rollouts_per_example. Automated eval runs read this config when no sample count is supplied; infinite tasksets default to 50 examples.

^{Macroscope summarized 6476f69.}