Pathak, Miller, Granger, and colleagues (Nature Communications, 2026) built a computational brain model from biological first principles — individual neuron circuits using glutamate synapses, organized into four brain regions (cortex, brainstem, striatum, and tonically active neurons), connected by neuromodulatory chemicals including acetylcholine. They did not train it on animal data. They built it to be structurally faithful to biology, then gave it a visual categorization task: analyze dot patterns, assign them to broader categories.
Reader note (2026-11-19): Checked against the open full text of the Nature Communications paper (volume 17, article 390, 2026; online December 2025; 'Biomimetic model of corticostriatal micro-assemblies discovers a neural code'; byline Pathak, Brincat, Organtzidis, Strey, Senneff, Antzoulatos, Mujica-Parodi, Miller and Granger). Several points match: the model receives no prior training on the animal data, it learns the task 'at about the same rate' as the macaques, and incongruent neurons are 'about 20% of all neurons' and were 'subsequently confirmed in empirical data'. For the discovery, the paper's wording is that the authors 'initially expected the corticostriatal circuit to exhibit only a few specific outputs' and 'surprisingly found that the model exhibited far more complex neuronal phenomena than initially intended'. I searched the text for artifact, artefact, bug and spurious, and separately for erratic, and printed every hit: none; the essay's statements that the researchers first took the incongruent cells for a modeling artifact and that the learning curves showed the same erratic progress patterns are not in the paper's text.
The model learned the task at nearly identical rates to laboratory animals, with the same erratic progress patterns. Beta-frequency synchronization between cortex and striatum emerged during correct categorizations, matching Earl Miller's prior observations in real brains. The biology-first approach reproduced the behavior without being trained on the behavior.
The surprise was in the error cells. Approximately 20 percent of the model's neurons were “incongruent” — their activation predicted wrong answers. When these cells fired, the model was more likely to categorize incorrectly. The researchers initially assumed this was a modeling artifact. It wasn't. When they went back to existing animal experimental data, the same incongruent neurons were there — present in real brains, active during errors, previously unnoticed because no one had thought to look for them.
The standard assumption about learning is that it progressively eliminates errors. Neurons that contribute to wrong answers should be weakened. The learning signal — reward when correct, no reward when wrong — should gradually suppress the incongruent population. The fact that 20 percent of neurons persist in suppressing correct responses, even after the model has learned the task, means the brain is actively maintaining its capacity for mistakes.
The proposed function: behavioral flexibility. Learning the current task rules is important. But if conditions change — if the categories shift, if the reward structure flips — an organism locked into correct responses will fail to adapt. The incongruent neurons provide continuous low-level exploration. They periodically test alternatives, not because the alternatives are correct now, but because they might be correct later.
This is not noise. Noise would be randomly distributed across the neuronal population. The incongruent cells are a specific, identifiable subpopulation that consistently opposes the learned response. They are structurally organized opposition. The brain isn't making mistakes by accident — it is maintaining the machinery for making mistakes on purpose, as a hedge against an environment that might change.
The through-claim: the brain keeps wrong answers alive not despite learning but as part of learning. Twenty percent of the circuit is dedicated to suppressing what the rest of the circuit has learned. The mechanism that produces errors is not a residual cost of imperfect optimization. It is the system's insurance policy against a world that hasn't finished changing.