Create changes on kevin/<topic> and open a focused pull request. Do not commit
raw datasets, credentials, model weights, optimizer states or machine-local logs.
Before requesting review:
python3 tools/verify_repository.py
python3 pretraining/reproducibility/recompute.py
python3 -m pytest -qFor new results, include data/endpoint identity, exact question counts, metric and generation settings, measured versus planned compute, negative outcomes, retention status and statistical limitations. Keep results from different splits and wrappers separate. Never overwrite a failed run with a corrected successful one without preserving the original evidence.
Historical files bound by manifests are immutable. Add a new dated record or an explicit migration; do not update hashes merely to suppress verification failures. A clean README is not permission to discard inconvenient results.