Experiment notebook / Reading notes
Compare experiments fairly
Prevent environment and input changes from obscuring the question.
Keep a baseline
Choose a reference run with a clear reason. Compare the same input split and objective rather than selecting whichever historical result looks most favourable.
Change one dimension
Separate code, configuration and dataset changes when possible. If several must change together, state that the comparison cannot isolate their individual effects.
Report variation
Repeat stochastic work when the decision depends on small differences. Record random seeds while acknowledging that a seed alone may not guarantee identical execution.