Experiment tracking¶
Maturity: verified
Experiment tracking owns the durable identity, lineage, metrics, artifacts, and terminal status needed to inspect and compare work across a logical run and its execution attempts. A run is the stable experiment identity; each attempt is a distinct execution that contributes evidence without overwriting another attempt.
Attempt evidence¶
The current bounded application implements tracking as attempt-local files on local or mounted storage. It does not integrate TensorBoard, MLflow, Weights & Biases, or a tracker plugin.
Evidence set¶
One successful attempt publishes:
the complete resolved YAML config in a content-named file;
metrics.jsonl, containing one record per completed optimizer step;scheduled PyTorch checkpoints;
run_manifest.json, published last as the terminal inventory.
The manifest records run, attempt, config and training-contract identities; dataset name, snapshot, split and split fingerprint; selected implementations; device and optional provider provenance; resume lineage; status and final step; and byte counts plus SHA-256 hashes for referenced artifacts.
On a non-finite training error, the application also writes
non_finite_training_error.json. The failed manifest identifies the failing
stage and attempted step, and includes only checkpoints completed by the
current attempt.
What metrics mean¶
The current step record contains event_type="optimizer_step", global_step,
loss, num_target_patches, and the ordered sample_ids. It is completed-step
evidence, not a validation or benchmark result.
Provider provenance¶
Managed callers may supply the exact source revision, immutable container image URI, provider job ID, launch-config URI, and launch-config SHA-256. The application records these values but does not independently attest that the provider ran them.