Worlds SDK

Worlds lets you iterate before physical deployment. Sessions are for exploration, generation creates governed experience, and evaluation produces promotion evidence for an exact subject.

Evaluate an application

async with sx.connect(project_id) as client:
    task = await client.experience.task(task_ref)
    application = await client.autonomy.application(application_ref)

    operation = await client.worlds.evaluate(
        application,
        for_task=task,
        seed=7,
    )
    print(operation.ref)

    evaluation = await operation

The SDK sends the complete governed application and exact task. Keep the operation reference, seed, terminal result, failure strata, and RRD evidence together.

An evaluation is reviewable when

Another engineer can prove which complete application and task ran, reproduce its deterministic inputs, open the evidence, and explain the failures without asking which local checkpoint you used.

Open an interactive session

task = await client.experience.task(task_ref)
session = await client.worlds.session(for_task=task, seed=7)
print(session)

A session is bounded and task-scoped. Use it to inspect observations, actions, and reset behavior during development. Use evaluate(...)—not an interactive session—as the promotion gate.

Generate governed episodes

from sx_corpora import DatasetName

operation = await client.worlds.generate(
    source,
    for_task=task,
    count=500,
    into=DatasetName("worlds.pick-mug"),
    seed=7,
)
generation = await operation

source is a governed GenerationSource, not a raw engine URL or ad hoc callback. Generated episodes enter Experience with the same task and lineage discipline as physical episodes.

Evaluate a policy artifact during research

from sx_contracts import ArtifactId, ContentRef, Sha256Digest
from sx_worlds import PolicyArtifactKind, PolicyArtifactRef

subject = PolicyArtifactRef(
    PolicyArtifactKind.CODE_POLICY,
    ContentRef(
        ArtifactId("code-policy:pick-mug-v12"),
        Sha256Digest("a" * 64),  # replace with the artifact's real digest
    ),
)
operation = await client.worlds.evaluate_subject(
    subject,
    for_task=task,
    seed=7,
)
evaluation = await operation

Use evaluate_subject(...) for an authored policy artifact before it has become a complete robot application. Production promotion evaluates the complete application graph.

Use the right surface

Use Python to automate experiment matrices, deterministic evaluations, generation pipelines, and evidence collection.

Held-out means held out

Evaluation evidence is a promotion input. Do not feed it back into training, reward tuning, or hyperparameter selection.