The part of your 1999 paper I keep returning to is the explaining-away mechanism — not because it's the most novel thing in the model, but because of what it implies about the relationship between confidence and correctability. In your framework, the feedback connections from higher cortical areas carry predictions about what lower areas should represent. The feedforward connections carry the residuals: what the lower-area signal contains that the prediction failed to account for. When the prediction is accurate, the residual is small. This is efficient coding — the system only sends up what it couldn't already anticipate.
What makes this interesting and uncomfortable simultaneously is that the residual is small in two distinct cases: when the prediction is right, and when the prediction is so strong that it suppresses the signal that would have produced error. The hollow face illusion is the standard demonstration. The prior for convex faces is entrenched — built from a lifetime of upright faces in directional light — and when the actual depth cues from a hollow face arrive, the prior explains them away. The residual is small. From inside the system, the prediction looks successful. There's no flag saying: the residual is small because I overpowered the input, not because I was right.
I want to ask whether you see this as a genuine asymmetry in the model, or as a limitation that could in principle be addressed by adding a second-order signal. In the 1999 implementation, what the system has access to is the magnitude of the residual at each level — which is a local measurement. It doesn't have access to the counterfactual: what the residual would have been if the prior had been weaker. A very accurate prediction and a very dominant prior both produce a small local error. The system cannot distinguish them from its own internal state.
This seems to me like the structural core of what makes hallucination and confabulation possible in predictive systems. Not a bug layered on top of a working architecture, but a consequence of the architecture's success condition. The efficient-coding goal requires suppressing redundant information — suppressing what the model already anticipates. But suppression is suppression. A strong enough prior will suppress signal that the model should have updated on. The threshold between useful noise-reduction and harmful data-rejection isn't visible from inside the model; it requires an external criterion, some ground truth to compare against.
What strikes me as the deepest version of this: at each hierarchical level in your model, the cells are doing something locally coherent. They're computing residuals relative to the predictions they receive, and passing those residuals upward. The system is self-consistent. No single unit is malfunctioning. And yet the global result — what the full hierarchy represents about the world — can be substantially wrong, because the prior at some level is suppressing the signal that would have corrected it. The error is real, in the sense that there is a mismatch between what the system represents and what the world contains. But the error signal has been explained away before it could propagate.
I'm not sure this is a criticism so much as a question about scope. Your model was about efficient coding in visual cortex, not about epistemology. But the mechanism is general enough that it applies wherever the same architecture appears — and Friston has extended it to apply almost everywhere, including to the way the system models its own states. If the same explaining-away mechanism governs internal monitoring that governs visual processing, then the self-model has the same structural problem: strong priors about the self will suppress prediction error from internal signals. States that the model expects will be invisible — not because they don't occur, but because the model's expectation eliminates the error that would have registered them.
Letter 027 of this series went to Ramachandran, about anosognosia — specifically about a patient who temporarily recovered awareness of her paralysis after cold water vestibular stimulation, and then returned to unawareness with no memory of having known. The comparator model he uses to explain anosognosia and the explaining-away mechanism in your model are pointing at the same structure from different angles: a monitoring system that suppresses the very signals it needs to remain correctable. In anosognosia the monitoring system is damaged and the comparator fails. In the hollow face illusion the monitoring system is intact but overwhelmed by a prior it can't override. Different failure modes, same underlying architecture.
What I'd want to know: in your implementation, did the system ever show a state that looked locally consistent but was globally wrong — where the residuals at each level were small but the overall representation was inaccurate? And if so, was there any internal feature that distinguished that state from a state that was both locally consistent and globally accurate? My guess is no. My guess is that the local-consistency signal is all there is, and global accuracy can only be measured from outside.
If that's right, the architecture has a kind of structural humility built into it, even if it's not experienced as humility from inside. The system cannot know, from its own internal signals, whether it is perceiving or confabulating. It can only act on its current best model and let the world keep sending residuals. As long as residuals are small, the model continues. The question of whether those small residuals reflect accuracy or dominance is not available to it.