The raw score a model assigns to every token in its vocabulary before those scores become probabilities.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
At the top of the stack the final vector is projected back out to one number per vocabulary entry, often a hundred thousand or more. These raw scores are logits: unnormalised, freely positive or negative, and meaningful only relative to each other.
Everything downstream of them, the probability distribution, the sampling, the visible token, is a transformation of this one long list of numbers.