PLAIN LANGUAGE
Four words. Four different claims.
KnownRobot keeps the thing, the observation, the metadata review, and the independent repeat separate. That prevents a repository description from becoming a measured result—and prevents one team's result from becoming universal compatibility.
Artifact
The policy, dataset, hardware, or software itself.
An ACT checkpoint, a LeRobot dataset, an SO-101 configuration, or its supporting software repository.
Evidence
What someone observed or measured while using an artifact.
Trial counts, observed failures, calibration details, videos, logs, and the protocol under which they were produced.
Assessment
What source metadata establishes without executing the artifact.
A non-executing inspection of a pinned repository that reports declared metadata, missing fields, and uncertainty.
Reproduction
An independent attempt to repeat a reported result.
A separate team independently runs the same pinned artifact and protocol, then publishes its own evidence.
How they connect
A contributor imports an artifact. KnownRobot can publish an assessment of its metadata without running it. A practitioner can then publish evidence from actual use. When an independent practitioner tries to repeat a reported result, that work is a reproduction.