A research notebook that your future self can use

A useful research notebook does more than store successful results. It preserves enough reasoning to explain why an experiment was run and what should happen next.

The format can be simple. A text file, a versioned document, or a structured notebook can work if the record is consistent and easy to find.

One question per entry

Start with the question the experiment is meant to answer. Keep it narrow enough that the result can guide a decision. “Try a different model” describes an activity; “Does this change reduce a particular failure case under the same evaluation?” describes a question.

Then write down what stays fixed. This makes it easier to tell whether a later comparison is actually measuring the intended change.

Record the recipe

Save the data version, code revision, configuration, environment information, and the command or procedure used. Record random seeds where they apply, while recognizing that a seed alone does not guarantee identical behavior across all environments.

Keep large outputs in an appropriate storage location and reference them from the entry. A notebook should point to the evidence rather than become an unsearchable collection of screenshots.

Separate observation from interpretation

Write what happened before explaining what it might mean. A run failed, a metric changed, or a pattern appeared in a set of examples: these are observations. An explanation is a hypothesis to investigate, not automatically a conclusion.

A short entry can use this structure:

1. Question: What uncertainty is this experiment addressing?
2. Setup: What data, configuration, and procedure were used?
3. Observation: What happened, with links to the outputs?
4. Interpretation: What explanations are consistent with the evidence?
5. Next step: What small test would distinguish those explanations?

Keep negative results

An unsuccessful approach can still explain a constraint or prevent repeated work. Record it with the same care as a promising result, especially if it changed the direction of the project.

A good notebook reduces the cost of returning to a problem after time away. It gives future collaborators, and your future self, a path from the question to the evidence.