Reproducibility¶
Maturity: verified
Reproducibility requires recording the inputs, identities, lineage, and runtime choices needed to understand how a result was produced and what can be repeated.
Reproducibility scope¶
The current bounded path provides strong identity and lineage evidence. It does not promise bit-identical reruns or exact general-training resume.
Recorded identities¶
The application records:
full resolved-config and resume-compatible training-contract hashes;
logical run ID and distinct attempt ID;
canonical dataset name, snapshot ID, split, and split fingerprint;
selected model, objective, and optimizer implementation keys;
checkpoint parent URI and producing attempt;
optional source, container, provider-job, and launch-config identities;
per-artifact SHA-256 hashes and byte counts in the terminal manifest.
Determinism controls¶
The bounded runner calls torch.manual_seed(config.seed) before model
construction, uses an ordered canonical split, sets shuffle=False, and
restarts the loader at its beginning if more bounded steps are requested than
one pass supplies.
Reproducibility limitations¶
The current contract does not persist random-number-generator state, loader or sampler position, scheduler or mixed-precision state, worker seed derivation, backend determinism settings, package lock identity in the attempt manifest, or distributed rank state. A resume restores model and optimizer state at a known completed step but begins a fresh loader iterator.
Treat the accepted one-step and resume-to-step-two evidence as an integration proof for the bounded application. The current application does not support broader train/validation or exact-resume claims.