Experiment notebook / Reading notes
Design for interrupted runs
Checkpoints and partial results that are safe to continue.
Choose a checkpoint boundary
Save a consistent unit of progress after the required data has been written. A file existing on disk does not necessarily mean a checkpoint is complete.
Validate before resuming
Check input versions, code compatibility and checkpoint integrity. Reject an incompatible continuation instead of silently combining unrelated execution states.
Retain the explanation
Link a resumed run to its predecessor and record what changed. Preserve the interruption reason and distinguish recovered outputs from the original partial result.