Cognitive modeling rarely involves hypothetico-deductive reasoning

It involves formalising and updating beliefs about cognitive processes

It can require intuition, and may rely on exploiting patterns in data to build theory

This is still science. Here's an e.g. of why...

The panels in the above plot present predicted-vs-real contributions on an economic game for three models.

Bottom panels plot distribution of errors.

Model 1 has more error, but is less biased. Model 2 has less error, but systematically underestimates. Model 3 has best of both

Model 3 reduces both the error and the bias in predictions made by older models. This is real progress. Great.

Now look more closely at the plot. The detail may afford more progress

See those weird bands? What are they? And why are they in models 2 and 3, but not in model 1?

Note that the bands are at the values 5, 10, 15, and 20. How can we explain them? Here's where common sense comes in.

In economic games, certain contributions are more likely. People round their contributions. They'll give 15 tokens instead of 16, or 10 instead of 12.

So the weird bands in the plot for models 2 and 3 tell us something. The model does not explain what happens when people round their contributions. Whatever process the model is formalising, it isn't capturing something about this rounding. We have an explanatory gap.

Their absence in model 1 also tells us something. This model implements a pure reinforcement learning model. This suggests that whatever mechanism might be involved in determining how people round their contributions in economic games, it might be related to learning processes.

So now we can go back to the literature, and think about how people set the granularity/rounding of contributions in economic games, and whether learning might be involved, and make more progress with the model.

Important point is the thought process. We can now build a new model and pre-register its predictions and test it on new data.

That might make the model more believable. But that's not where the progress happened (assuming success) Progress is in the model building process

Purpose of this thread is not to criticize pre-reg or anything. Simply that I've seen a few people describe this process lately, and I thought this seemed like an example that might get the intuition across