The wiring diagram is a possibility map

Imagine receiving the complete road map of a city but no traffic lights, opening hours, weather, vehicle positions or destinations. The map constrains every journey, yet it does not tell you which route will carry traffic at 08:45. A connectome is similar. It tells us where influence could travel. It does not, by itself, tell us where influence is travelling now.

Classical descriptions of cortical hierarchy were never as simple as a conveyor belt. Anatomical work mapped visual and other cortical areas into distributed levels using laminar patterns of feedforward and feedback projections.[1] Later tracer studies found a dense, directed and strongly weighted interareal network, including many pathways that earlier methods had missed.[2] The result is a hierarchy with direction and asymmetry, but also recurrence, shortcuts and loops.

That matters because a hierarchy can order constraints without dictating a single path. Early sensory areas tend to supply fine-grained evidence. Higher areas can pool across larger spatial and temporal contexts. Feedback can then change what earlier populations amplify, suppress or complete. The hierarchy supplies different receptive fields, delays and projection patterns. Recurrent traffic decides which of those resources matter for the present task.

Structural topology also shapes spontaneous dynamics before a task begins. Simulations built on empirical cortical connectivity have reproduced aspects of functional coupling across multiple timescales, showing that the scaffold biases which patterns are easy to sustain and which are difficult to reach.[3] Yet the mapping is many-to-many. One anatomical graph can support several dynamical regimes, and similar correlations can arise from different combinations of delays, local dynamics and input. Structure therefore acts more like a set of valleys and ridges than a fixed itinerary.

This resolves an apparent contradiction. Brain architecture can be stable enough to preserve a person’s characteristic organisation while remaining flexible enough to reconfigure during a decision. The stable part supplies the feasible state space. The changing part selects a region of that space. Flexibility is not freedom from anatomy; it is controlled movement within anatomical constraints.

Three maps that should not be collapsed

Neuroscience uses “connectivity” for several different objects. Collapsing them creates the illusion that a single graph can answer every question. It cannot.

Structural connectivity is the possibility map. Functional connectivity is a pattern in observations. Effective connectivity is a claim about influence. A functional edge can appear without a direct pathway because two areas share an input. A structural edge can exist without measurable functional coupling because its signals arrive at an ineffective time. An effective edge can also be conditional: present during attention, absent during distraction, reversed under inhibition.

Even within one connectivity type, the graph depends on scale. A cortical column, parcel, lobe and whole brain are different nodes. Averaging thousands of neurons into one parcel can hide opposing subpopulations. A one-second estimate can merge several gamma cycles; a one-minute estimate can merge several cognitive episodes. The question “did the network change?” is therefore incomplete until the node definition, edge definition and temporal grain are specified.

A useful discipline is to name the physical event behind every symbol. In the four-node example below, a thicker edge does not mean more axons appeared. It means a fixed signal arriving from A2 produces a larger change in B1 under the stated state. If the model cannot say what would be held constant and what would be measured, its dynamic graph is only a picture.

A compact causal model

We can compress the intuition without pretending the compression is the brain. Let S_ij be the structural capacity from source j to target i. Let C_ij(t) represent how available that route is at time t because of relative timing, phase and local circuit state. Let g_i(t) represent the target population’s gain, meaning how strongly its output changes for a given input change. Then a minimal effective-weight model is:

Population activity can then be written schematically as x(t+1) = φ(E(t)x(t) + u(t)), where x is current activity, u is external input and φ is a nonlinear response. The point is not the equation’s elegance. The point is the intervention it suggests. Hold S fixed. Change C or g. If downstream responses change in the predicted direction, the network has reconfigured functionally even though its anatomical graph has not.

How the scaffold breathes

A fixed scaffold becomes many usable networks through three coupled operations: signals return, windows open and close, and populations change their susceptibility. Recurrence, timing and gain are therefore best understood as route-making mechanisms.

Recurrence turns a sweep into a negotiation

A feedforward sweep can classify many clear images quickly. It is less reliable when evidence is occluded, cluttered or locally ambiguous. Under those conditions, later populations can send constraints back towards earlier representations. The returning signal does not carry a finished picture. It changes which local interpretation remains stable.

Experiments in primate object recognition found images for which behavioural solutions and late inferotemporal responses were poorly explained by feedforward models. The timing and perturbation evidence was consistent with recurrent circuits contributing to the solution.[11] This does not imply that every recognition event requires the same recurrence. It implies that the effective architecture depends on the problem’s demand.

Oscillations create recurring windows

Neural populations do not respond equally at every instant. Excitability rises and falls as local circuits oscillate. A signal arriving near a receptive phase can have a larger effect than an equally strong signal arriving near a suppressive phase. This makes timing a candidate routing variable.

The communication-through-coherence framework proposes that phase relations can make communication selective by aligning output from one population with high-gain periods in another.[8] The framework has generated valuable tests, but coherence is not self-interpreting. A common input can synchronise two areas. Strong communication can itself generate coherence. Oscillatory power and anatomical coupling can create interareal coherence without coherence being the independent cause.[9]

Frequency labels should not be treated as fixed message types. Theta, alpha, beta and gamma describe ranges in a spectrum, not universal semantic channels. Their functional role depends on circuit, species, task, layer and measurement. In a large intracranial memory study, theta networks tended to synchronise while gamma networks tended to desynchronise during successful encoding and retrieval, with regional gamma power related to theta synchrony.[13] The result argues against the simple slogan that more synchrony at every frequency means more integration.

Oscillations can support routing in at least three distinct ways. They can create periodic windows of excitability. They can coordinate when spikes from separate sources arrive. They can organise slower and faster processes through cross-frequency relations. These possibilities produce different interventions. Phase-shifting a receiver tests the first. Changing interareal delay tests the second. Selectively disrupting phase-amplitude coupling tests the third. A spectral peak by itself distinguishes none of them.

Neuromodulation changes the slope of response

Timing alone cannot explain why the same input sometimes produces a small local correction and sometimes recruits a broad system. Neural gain supplies another control surface. In a simple input-output curve, higher gain means a small input difference produces a larger output difference near the operating point. In a recurrent network, that local change can alter which attractor wins, how far activity spreads and whether modules remain segregated.

Large-scale modelling has shown that increasing neural gain can move a structurally fixed network towards more integrated dynamics.[6] Human task-fMRI studies have reported more integrated network states during better task performance, with associations to pupil dilation, a proxy related to arousal systems rather than a direct neurotransmitter measurement.[5] A later analysis across seven tasks found low-dimensional trajectories whose spatial axes aligned with regional neuromodulatory receptor-density patterns and structural controllability estimates.[7]

These findings support a plausible bridge from neuromodulatory systems to changing network topology, but the evidential types must remain separate. The gain study is a computational demonstration of capacity. The pupil result is an observational association. Receptor-density alignment is spatial correspondence. None alone proves that one transmitter set a specific effective edge during one decision.

Neuromodulation is also neither unitary nor purely global. Noradrenaline, acetylcholine, dopamine, serotonin and other systems differ in receptors, projection patterns, timescales and cellular effects. A diffuse projection can still have region-specific consequences because receptor densities and local circuits differ. Work linking ascending arousal nuclei to cortical state landscapes supports the idea that gain can reshape which global states are reachable, but it also shows why “arousal” cannot be used as a single explanatory variable.[17]

Gain has a cost curve. Too little makes weak evidence ineffective. Moderate gain can sharpen selection and enable distant modules to influence one another. Too much can compress distinctions through saturation, amplify noise or trap the network in a dominant state. This produces a testable prediction: performance should depend on the match between gain and task demand, not on gain alone. A task requiring stable local processing may worsen under the same gain increase that helps a task requiring rapid integration.

The integration-segregation balance

Local modules protect specialised computation from interference. Cross-module integration allows evidence, goals and memory to constrain one another. A useful brain therefore needs neither maximal isolation nor permanent global coupling. It needs controlled movement between regimes.

Learning studies have observed time-varying modular organisation as participants acquire a motor skill.[4] Other work has related integrated states to task performance, while carefully sampled data show that stable group and individual organisation still explains more network variance than task or day-to-day state.[15] These results are compatible. The scaffold and personal network fingerprint can dominate overall variance while smaller state-dependent changes still carry causal importance.

flowchart TD
  A["The wiring diagram is a<br/>possibility map"] --> B["How the scaffold breathes"]
  B --> C["When a changing network is<br/>measured badly"]
  C --> D["Run the gated-scaffold test"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#5B5CF6,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

When a changing network is measured badly

A dynamic graph can reflect a changing brain, a changing estimator or both. Before interpreting a “network state”, ask what edge was measured, at what temporal grain and which rival process could create the same pattern.

Correlation can impersonate routing

Suppose activity in region A predicts activity in region B. One story says A sends information to B. Another says region C drives both. A third says B drives A. A fourth says slow haemodynamic filtering makes independent neural events overlap. Functional connectivity alone does not choose among them.

Model-based effective-connectivity methods can narrow the possibilities, but they do not escape assumptions. Granger-style prediction depends on temporal sampling and omitted variables. Dynamic causal models depend on a specified generative model and candidate architecture. Perturbational readbacks depend on stimulation spread and the observation process. The remedy is not to reject models. It is to expose the assumptions that turn data into an edge and to compare rival models that could produce the same observation.

The strongest design combines paths that fail differently. Anatomy limits candidate routes. Electrophysiology resolves timing. Perturbation tests influence. Behaviour tests relevance. A mechanism becomes more credible when all four converge and when the same manipulation leaves a matched control route unchanged.

Synchrony carries the same ambiguity. It may increase communication because excitable phases align. It may be produced by communication. It may arise because two areas follow a shared rhythm. This is why the 2025 theta-stimulation experiment is informative: it changed phase targeting while holding the approximate stimulation rate constant, then measured evoked responses as well as synchrony.[10] The design moves beyond correlation, though its small clinical sample and specific hippocampal network constrain transfer.

The window can manufacture a state

Time-resolved fMRI often estimates connectivity inside sliding windows. A short window reacts quickly but has few samples, so correlations become noisy. A long window is statistically steadier but blurs transitions. Simulations and empirical analyses have shown that common windowed methods can have poor power or misleading behaviour unless tested against suitable null models.[14] The broader field now treats dynamic functional connectivity as a family of estimands rather than a single fact.[16]

Sampling rate sets another boundary. BOLD measurements unfold over seconds and are filtered through vascular responses. They can reveal slow reorganisation across tasks or sustained states, but they cannot directly resolve millisecond phase gates. EEG and MEG improve temporal resolution while mixing sources spatially. Intracranial recordings offer local precision in selected clinical locations. No modality supplies the whole network without an inference bridge.

Stable network organisation is not an embarrassment for dynamic accounts. Highly sampled individuals show that shared and person-specific organisation dominates over task and daily variation.[15] The correct prediction is therefore not that the brain becomes a different graph every second. It is that small, structured changes in effective influence occur on top of a strong baseline and that those changes matter disproportionately when they control a bottleneck, hub or recurrent loop.

What would count as stronger evidence?

The best evidence depends on the claim. A claim about anatomy needs pathway measurement. A claim about temporal coordination needs frequency-specific, time-resolved measurement and phase controls. A claim about causal routing needs perturbation and downstream readback. A claim about neuromodulatory control needs manipulation or a converging bridge from receptor distribution, physiological state and network dynamics.

The strongest alternative explanation

A critic may say that architecture did not change at all. Only node activity changed inside a stable network. This objection is correct if “architecture” is reserved for anatomy. It is too narrow if the scientific question concerns conditional causal organisation. In control systems, the same components and wires can implement different closed-loop dynamics when gains and feedback paths change. In brains, effective architecture names the current pattern of influence, not a second hidden anatomy.

This definition also prevents overreach. We should not infer literal rewiring from a changing correlation matrix. We should not infer a new cognitive module because a clustering algorithm drew a boundary. We should say exactly which level changed: activity, statistical dependence, estimated influence or physical connection.

Run the gated-scaffold test

The executable lab below holds structural connectivity fixed while varying cross-module alignment, target gain and recurrent strength. It exposes the causal logic that prose can hide.

The contribution: a Gated Scaffold Test

The proposed test asks whether an apparent network change is better explained by a changed pathway, a changed timing gate, a changed gain regime or a measurement artefact. Its smallest useful form needs two modules, one weak cross-module structural route, a controllable phase or timing relation, a gain parameter and a downstream readback.

The protocol has four stages. Calibrate the observation model on synthetic stationary and known-transition data, so the estimator’s false states and temporal blur are known. Constrain the route with anatomy and a matched alternative pathway. Intervene on one gate while matching source energy, trial count and behavioural context. Read back both local state and downstream response, so a routing effect can be separated from a general excitability change.

Pre-registration should name the primary interaction. For example: phase-targeted stimulation will increase the downstream evoked response more than phase-blind stimulation, and the difference will be larger at moderate target gain than at low gain. This interaction is more discriminating than a main effect of stimulation because shared drive, power and arousal can produce main effects without the proposed gate.

How to read the lab

Start with alignment at 0.55, gain at 1.20 and recurrence at 0.75. Run the model. Module A receives the same twelve-step input on every trial. Now press “Phase-scramble”. Cross-module alignment drops while anatomy and average input remain fixed. If Module B’s peak falls, the simulation demonstrates timing-sensitive routing. Raise gain and run again. If B recovers, the model shows how target susceptibility can compensate for a weak timing gate, within stability limits.

The cross-edge share is the proportion of total effective weight that spans the two modules. The B/A propagation metric compares the peak response of Module B with Module A. Late-state load is the mean squared activity over the final twenty steps. It is a crude warning signal for activity that persists or saturates after the input has ended. These are transparent synthetic measures, not clinical biomarkers or accepted universal network metrics.

A positive result permits a narrow conclusion: under the tested scaffold and operating range, changing timing or gain changed downstream influence while the structural matrix remained fixed. It does not permit the conclusion that the same gate explains a human cognitive state. A negative result says the chosen gate, range or readback failed to account for propagation. It may weaken the hypothesis, or it may reveal that another pathway, delay or nonlinearity dominates.

The artefact cannot establish consciousness, representation, intention or the biological identity of a neuromodulator. It cannot tell whether two real brain areas communicate through direct axons or an unmodelled intermediary. What it can do is force the researcher to state the causal variables, intervention and prohibited inference before collecting impressive-looking networks.

The proposal should not be applied where the temporal resolution is too low to resolve the claimed gate, where anatomy is unknown enough that alternative pathways dominate, or where interventions change source drive, timing and gain simultaneously. In those cases, the model is a vocabulary for uncertainty, not a mechanism result.

from __future__ import annotations

import json
import math
import random
from dataclasses import dataclass


@dataclass(frozen=True)
class Config:
    alignment: float = 0.55
    gain: float = 1.20
    recurrence: float = 0.75
    steps: int = 80
    seed: int = 7


def structural_matrix() -> list[list[float]]:
    """Return one fixed, directed, two-module scaffold."""
    n = 8
    s = [[0.0 for _ in range(n)] for _ in range(n)]
    for module in (range(0, 4), range(4, 8)):
        for i in module:
            for j in module:
                if i != j:
                    s[i][j] = 0.12 + 0.04 * ((i + 2 * j) % 4)
    for i, j, weight in ((4, 3, 0.34), (5, 2, 0.238), (2, 5, 0.187), (3, 4, 0.272)):
        s[i][j] = weight
    return s


def simulate(cfg: Config) -> dict[str, float | list[float]]:
    rng = random.Random(cfg.seed)
    s = structural_matrix()
    n = len(s)
    x = [0.0] * n
    trace_a, trace_b = [], []

    effective = [[0.0 for _ in range(n)] for _ in range(n)]
    for i in range(n):
        for j in range(n):
            same_module = (i < 4) == (j < 4)
            timing = 0.92 if same_module else cfg.alignment
            target_gain = 1.0 if i < 4 else cfg.gain
            effective[i][j] = s[i][j] * timing * target_gain

    for t in range(cfg.steps):
        external = [0.0] * n
        if t < 12:
            external[0] = 0.85
        next_x = [0.0] * n
        for i in range(n):
            recurrent_input = sum(effective[i][j] * x[j] for j in range(n))
            noise = rng.uniform(-0.005, 0.005)
            drive = 0.45 * x[i] + cfg.recurrence * recurrent_input + external[i] + noise
            next_x[i] = math.tanh(drive)
        x = next_x
        trace_a.append(sum(x[:4]) / 4.0)
        trace_b.append(sum(x[4:]) / 4.0)

    total = sum(sum(row) for row in effective)
    cross = sum(
        effective[i][j]
        for i in range(n)
        for j in range(n)
        if (i < 4) != (j < 4)
    )
    peak_a = max(abs(v) for v in trace_a) or 1e-9
    peak_b = max(abs(v) for v in trace_b)
    late = trace_a[-20:] + trace_b[-20:]
    late_load = sum(v * v for v in late) / len(late)

    return {
        "cross_edge_share": cross / total,
        "b_over_a_peak": peak_b / peak_a,
        "late_state_load": late_load,
        "module_a": trace_a,
        "module_b": trace_b,
    }


if __name__ == "__main__":
    result = simulate(Config())
    print(json.dumps(result, indent=2))

stateDiagram-v2
  [*] --> Claim
  state "Three maps that should not be<br/>collapsed" as Claim
  state "A compact causal model" as Test
  state "Recurrence turns a sweep into<br/>a negotiation" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Look for the changed gate before inventing a module

Stop treating a brain region’s activation or a correlation edge as the architecture of a computation. Ask which existing route became more causally effective, what opened it, and which intervention could close it again.

For experimental design, this changes the unit of explanation. “Area B activated during attention” is a location claim. “A phase-specific, gain-sensitive route from A through B changed downstream behaviour while source drive and anatomy were held fixed” is a mechanism claim. The second is harder, but it tells us what to perturb, what to reproduce and where a failure would live.

It also changes how null results are valued. If phase scrambling does nothing while a pathway lesion abolishes propagation, the route is structurally necessary but the proposed timing gate is not supported. If gain manipulation changes local firing but not downstream readback, the target became more excitable without network integration. These are informative separations, not failed stories.

For network analysis, it changes the order of operations. Begin with the biological timescale and causal question. Choose a sensor that can resolve them. Define what an edge means. Calibrate the estimator on stationary and known-transition simulations. Show that the apparent state survives reasonable analysis choices. Then intervene. Do not let graph construction silently decide what “network change” means.

For computational neuroscience and neuro-inspired engineering, it changes what should be copied. A static hierarchy of layers is a weak imitation of biological architecture. The more consequential design principle is bounded, state-dependent routing over a constrained scaffold: recurrent paths that can be interrupted, timing gates that can be tested, gain controls that can saturate, and readbacks that distinguish local activation from transmitted influence.

The brain’s wiring matters because it limits the possible. Its changing effective network matters because cognition happens in the selection among those possibilities. The practical decision is therefore simple: when behaviour changes faster than anatomy can, look for a changed gate before inventing a changed module.