Test assets

A fixture earns its place by being a pinned input with a known consumer, not by being realistic or popular.

Choose what the test needs

Start from the behavior under test, then select the smallest representative world, model, recording, dataset slice, checkpoint, or golden.
Prefer a recognizable public asset when it matches the behavior. It gives reviewers context and avoids maintaining an equivalent private fixture.
Add an asset only when an active test consumes it. Record the upstream revision, license, verification method, entrypoint, and derivation.
Reuse outputs from verified scenarios when useful: a generated map or bag can become the pinned input to the next layer.

Use a pinned pointer for large or independently maintained assets. Verify an immutable revision and checksum, and cache the fetched bytes.
Vendor only when the test must survive upstream availability or inspect the exact bytes, redistribution is allowed, and the project has a stated size budget.
Generate project-specific bags and goldens from a committed, seeded producer when no canonical source exists.
Compare re-simulated behavior with tolerance bands. Reserve byte equality for pure replay or deterministic transforms.

To select a known world, robot model, dataset, or recording, read references/canonical-assets.md and verify its dated evidence against the current source before adoption.
To define a project catalog, manifest, slice, or golden, read references/test-assets-layout.md.
Use examples/catalog.yaml only as a proven catalog shape, not as a universal asset set.
Use scripts/fetch_assets.py only for checksum-pinned pointer archives that match its supported manifest shape.
Testing owns the test layers and pass bars. Data owns training-data sourcing; the simulator and Hub skills own loading and transfer mechanics.

Every fixture has an active consumer, reproducible source or generator, license evidence, and an assertion whose tolerance matches how the fixture is produced.