More frustration today. I’ve been laying awake at night trying to go through mental checklists of all of the different details we might be getting wrong in our feature definitions. The worst part is that it’s possible nothing is wrong in the model itself. It could be that there’s a mistake in the model evaluation code, or really anywhere in the pipeline up to us finally reviewing results. Or maybe it’s all correct and our current feature set is bad. But does that mean we should spend more time re-checking things, or should we spend more time trying to come up with new things?