jupyter-python

Use nbclient with nbformat to execute notebooks programmatically. Read the notebook with nbformat.read, create a NotebookClient with an explicit kernel_name, timeout, and working directory, call execute(), then write the result to a separate output file with nbformat.write.

Restart the kernel and run all cells top-to-bottom in a clean process as the minimum reproducibility test. Set random seeds, control clock and network inputs, use parameter cells instead of scattered manual edits, and re-run from clean inputs to compare artifacts.

The kernel retains variables, imports, and state from earlier cells, so out-of-order interactive execution can hide missing definitions. A clean top-to-bottom run after restart exposes this hidden-state failure, which saved outputs and execution counts alone cannot prove absent.

Move reusable transformations, models, and validation into modules with unit tests whenever logic is reused or needs verification. Keep the notebook for orchestration, explanation, and display, and prefer a script or pipeline for unattended recurring execution.

Inspect outputs for credentials, personal row data, and large binary or base64 content, and verify error outputs and execution counts. Confirm the kernelspec and language metadata match the intended environment so recipients can reproduce the run.