What I want in an experiment note.

I want my experiment notes to be useful when I return to them a few weeks later. A result without the setup makes that difficult. For the PULSAR work, that means keeping the dataset, train/test split, seed, model settings and code version beside the score.

Write down the comparison

If I add a feature and the score improves, I need to know what I compared it with. The channel-only control in my multivariate experiments changed how I understood the results: some improvements over the first baseline were already there without the extra relationships.

Keep the explanation separate

An accuracy increase is a measurement. My explanation for it is a hypothesis. For example, channel relationships might help a handwriting dataset because movement across axes is coordinated, but the score alone doesn’t prove that explanation.

I’m trying to keep the next experiment small enough to answer one question. If I change the selector, the window sizes and the classifier together, it becomes harder to work out which change mattered.