Static ReadonlyDefaultThe default largest holdout-metric difference a retraining may show and still reproduce its run.
ProtectedFileWhere the snapshot files are read: deps.snapshotFiles, or the storage-provider store on the request's provider.
ProtectedLoadThe Evaluation Protocol a run was trained under, read through deps.protocols (default: RunView on deps.provider).
ProtectedLoadThe Prediction Unit version a snapshot was built from, read through deps.predictionUnits (default: RunView on deps.provider).
ProtectedReadReads the three files in parallel and refuses any whose SHA-256 differs from its column on the snapshot row (the
provenance file included: it is outside the content hash, so its own column is its only check), then refuses a
row whose verified matrix and manifest hashes and identity columns no longer yield its ContentHash.
Re-runs a completed unit run from its snapshot files and logged trainer request. The run and snapshot are read
through deps.snapshots and the files through deps.snapshotFiles (defaults: the RunView and storage-provider
stores on deps.provider, as deps.contextUser).
the MJ: ML Training Runs row to retrain
the training seams; only the snapshot stores, the trainer, the provider and the user are used
the largest holdout-metric difference that still counts as reproduced
both metric sets, their largest difference and whether the run reproduced
ProtectedRetrainSends a run's logged request settings to the trainer again, with new downloads of the verified snapshot files, under the protocol and unit the run was trained with; the retraining's holdout metrics.
ProtectedSnapshotWhere the run and its snapshot are read: deps.snapshots, or the RunView store on the request's provider.
ProtectedStoredThe run's results envelope, parsed once; a run without a readable envelope is refused, naming it.
StaticMaxThe largest |original − retrained| over every metric name on either side. A name present on one side only makes
the difference Infinity (the retraining did not measure what the run measured), and so does a non-finite value.
Retrains from Training Snapshots:
Retrainre-runs a completed run through the protocol trainer from its verified snapshot files, with its reads split into protected steps a subclass can replace.