Public API and commands

The small binding surface around native Inspect tasks.

Core Python API

These names are exported by inspect_labs:

Name	Purpose
`EnvironmentInfo`	Declare environment version, mode, capabilities and operations
`LabEnvironment`	Structural protocol for tools, read-only observation, artifacts and close
`LabEvidence`	Sample identity, environment and separately collected observation payload
`bind_task`	Attach environment setup, evidence scoring and cleanup to an unscored native task
`evidence_scorer`	Build a native scorer using collected laboratory evidence
`rescore_workflow`	Reapply a trusted judge to linked saved evidence with zero dispatch
`check_environment`	Check lifecycle, tools, observation stability and artifact mechanics
`ConformanceReport`	Report the clauses checked and any violations

bind_task

def bind_task(
    task,
    *,
    environment,
    judge,
    requires,
    evidence_dir,
    allow_physical=False,
    observation_timeout=30,
    metrics=("known", "correct"),
): ...

The factory receives native TaskState and returns a fresh LabEnvironment. requires is either a capability set or typed Requirements from inspect_labs.spec. metrics must include every key the judge can return, including known and correct. Existing task scorers are rejected to keep outcome ownership explicit.

rescore_workflow

def rescore_workflow(native_log, evidence_file, output, judge, metrics=None): ...

Paths are pathlib.Path objects. The output must not exist. Omitted metrics uses the keys recorded at binding time. Missing files or provenance mismatches fail clearly rather than rerunning the workflow.

Native execution

Use inspect eval for task execution and inspect view for native logs. The inspect-labs CLI is a convenience surface around native execution, discovery and replay, not another runner.

inspect-labs --help
inspect-labs list
inspect-labs doctor --environment litmus-measurement
inspect-labs run --task measurement --scripted
inspect-labs rescore run.eval --evidence run.labs --output replay.eval

run supports measurement and handoff reference tasks. Other tasks use native inspect eval. Live run requests require explicit --allow-live, a model and an enforceable --cost-limit. --price INPUT OUTPUT supplies missing model prices in USD per million tokens. A native per-sample limit is not a strict billing cap.

Upstream documentation

Inspect AI for tasks, models, agents, tools, approvals and logs
Inspect Robots for robot policy and embodiment interfaces
PyLabRobot for laboratory instrument interfaces

The API is provisional. See status before depending on it.