NumPy Random Seeds: Record the Experiment, Not Only the Seed

Define a small synthetic experiment

Create two fresh NumPy generators with np.random.default_rng(29) in the same local environment. Ask each for the same small sequence of integers using the same bounds, size and argument values. Save the code and both results. The exercise needs no real customer data and no model training.

It complements NumPy and analytical workflow foundations for Data Science learners in Jaipur. The purpose is to document a controlled comparison, not to promise identical results across every machine or future library version.

Compare fresh state with advanced state

Under the same generator, seed, calls, build and environment, the two fresh generators should produce the same stream. Now draw a second batch from the first generator without recreating it. Its internal state has advanced. Do not assume that reusing the same object restarts its earlier sequence.

Insert one extra draw before the recorded experiment and compare the later values. That change can shift what the experiment consumes even when the original seed remains unchanged.

Write a complete run note

Record Python and NumPy versions, the BitGenerator type, seed, method names, arguments and call order. Keep the dataset version and any ordering of input records as well. NumPy's stream-compatibility conditions are deliberately strict; default_rng does not provide a blanket guarantee of identical streams across all versions.

Separate repeatability from statistical quality

A repeatable random sample can still be poorly chosen for the question. Fixing a seed does not make the source data representative, remove leakage or prove that a model generalises. Review population coverage, data collection and evaluation design as separate questions.

Do not keep trying seeds until one produces a preferred result and then present only that run. If a larger study compares multiple seeds, document the planned method and report the relevant variation rather than hiding it.

Keep a compact handover

• A runnable script with a small bounded draw.
• The exact environment and random generator configuration.
• A note identifying any extra draw or altered call order.
• The separate question that this experiment does not establish.

Connect the record to future work

Another learner should be able to tell whether a mismatch comes from a changed input, environment or consumption order before debating a result. Mango Engineers' related Data Science learning path provides broader Python, statistics and workflow foundations; confirm current syllabus and prerequisites with the team.