nemo_gym.harness_capabilities.health

Scenario expectations over the public gym eval health-check engine.

Module Contents

Functions

Name	Description
`inspect_health`	Run health once on saved artifacts and compare each rollout’s check results.

API

nemo_gym.harness_capabilities.health.inspect_health(
    rollout_paths: collections.abc.Sequence[pathlib.Path],
    output: pathlib.Path,
    expectations: collections.abc.Mapping[str, nemo_gym.health.types.Verdict],
    steps: bool = True
) -> list[dict]

Run health once on saved artifacts and compare each rollout’s check results.

Registry entries default to healthy; scenarios explicitly declare exceptions. Task-level checks need repeated tasks and are outside this scenario contract. Missing records fail even when an unobserved health result was expected.