The only thing that can move

A stripped-down copy of a task is called a model. This model has one job: turn a simplified tasting record into one number, a price.

Two things affect that price.

The first is what the model is told about the bottle. Before it guesses, the tasting has already produced a description: cabernet present, plum strong, apricot absent. This description is the model's input.

Another tasting, even of the same bottle, could produce different inputs. But once this tasting's record has been supplied, those inputs stay fixed while the model prices it.

If the price is wrong, we do not let the error rewrite the tasting and pretend there was less plum or more apricot. That would change the evidence instead of teaching the model how to use it.

The second is how the model uses that evidence. It can change how strongly each input affects the price. This is the part we let the error adjust.

Suppose apricot counts for a lot. An apricotty wine then produces a high guess. If apricot counts for very little, the same input produces a lower guess.

Same wine. Same tasting. Different price, because the setting changed.

So this model has exactly one kind of adjustable setting: how much each input counts toward the price.

That is a design choice, not a claim that a human palate has only one way to change. Real learning can also change what a person notices and how clearly they tell sensations apart.

This model freezes all of that. It asks a narrower question: how much can be learned by changing only how strongly four inputs affect one answer?