"The moment the model encounters highly abstract, non-linear, or completely out-of-distribution reasoning (deeply nested code, intricate logical paradoxes, custom corporate APIs), the confidence gate shuts down."
It's funny to me you'd make this argument, because engrams make models much better at abstract reasoning, because the things responsible for reasoning don't have to reconstruct all the local features. The way you've framed this as "Embeddings aren't intelligence" is true, but it's also missing the point of what engrams do. They free up the parameters that are doing those things. Anyway, the actual measurements disagree with your assertions, and they're right their in the linked papers for you to actually research yourself.