Anchors model identity, versions, licenses and the access policy attached to each model, so a registry can be verified independently of its author.
Every compute job carries a precise model version anchor, so a receipt names exactly the weights that were run - not a moving target.
Signed evaluation summaries commit to the chain when useful; high-frequency benchmark detail stays off-chain and reviewable by reference.
A compact, replayable proof per job: model version, dataset slice, runtime build, cost and outcome - reviewable long after the run finishes.
Training and inference jobs declare their inputs, hardware class, cost ceiling and reproducibility bounds before execution starts.