The wiring itself reads the sequence: how axonal delay dispersion decides what a neuron detects, and what a myelin-free substrate computes by default

What the work claims

This is a theory paper with controlled simulation results, not an experimental study, and it says so. The central claim is that the set of axonal conduction delays arriving at a dendritic branch acts as a physical key: only input spike sequences whose pairwise timing differences happen to be cancelled by those delays land inside the branch's coincidence window, and a hard calcium threshold converts that synchrony into an all-or-none plateau. The novel quantity is the dispersion of the delay set, called Δ. The paper's three reported results: Δ moves a population of model neurons monotonically from event detection to order-selective sequence detection, with the transition emerging under random delays and random connectivity; Δ bounds the longest interval the code can represent and fixes an absolute timing tolerance of about a millisecond rather than a fixed percentage of conduction velocity; and that millisecond window multiplied by horizontal conduction velocity predicts cortical column diameter, with the two areas where horizontal conduction has been measured directly falling where the relation puts them.¹

The anatomical hook is myelination. Myelination raises conduction velocity by roughly an order of magnitude, and for a fixed spread of path lengths the delay dispersion falls as one over velocity, so whether a projection is myelinated sets the scale of Δ. If the framework is right, myelination is not merely a speed regulator but a switch on what the downstream neuron computes.

How it works

The motivation is the sparsity of cortical firing. The paper cites ferret V1 neurons responding to natural images at a median of about 4.1 spikes per second, fewer than 0.5 spikes per 100 ms perceptual window, and layer 2/3 pyramidal cells in somatosensory cortex firing spontaneously as slowly as about 0.1 Hz, with only about 5 percent responding to a tactile stimulus. At that sparsity a rate code carries almost nothing; timing must carry the information.¹

The model neuron, called an integrator neuron, has three domains. Distal dendrites receive the query, an upstream spike sequence whose information sits in firing order; each branch imposes its own set of delays, and where those delays compensate the departure-time differences of the sequence, spikes that left at different times arrive together. A coincidence window of about 1 ms, imposed in biology by fast feed-forward inhibition and in the model by a threshold on summed within-window arrivals, decides whether a dendritic calcium plateau fires. That plateau is hard-gated: below threshold nothing passes, above it the branch commits to a sustained current whose amplitude scales with match quality. The soma then checks the distal match against a proximal content signal, firing only when both agree. The canonical 1 ms window is anchored to direct measurements the paper reviews: a sub-2 ms integration window enforced by disynaptic feed-forward inhibition at hippocampal CA1 pyramidal cells, 1 to 2 ms for cerebellar Purkinje cells, and a local dendritic summation half-width of 10.8 plus or minus 0.5 ms in dentate granule cells that noradrenaline-enhanced inhibition narrows toward 4 ms.¹

In the simulations, populations of 1000 integrator neurons receive random projections from a 200-neuron feature population, of which two disjoint sets of 40 neurons form two events inside a gamma cycle. Conduction delays are drawn uniformly from zero to Δ, swept from 0 to 50 ms. At narrow dispersion, sequence detectors do not exist: 0.00 percent at Δ of 2 ms, against 9.76 percent of the population acting as event detectors. Sequence detectors first appear at Δ of 14 ms, peak at Δ of 18.6 ms at 0.180 percent of the population, and event detectors fall to 0.005 percent by Δ of 30 ms, with the two classes crossing at Δ of 16.4 ms. Crucially, the crossover dispersion scales with the inter-event interval with a fitted slope of 0.92 with a 95 percent bootstrap interval of 0.83 to 1.01, statistically indistinguishable from one: sequence detectors overtake event detectors when the delay spread is about equal to the interval being coded. Order selectivity is genuine, not conjunctive: pooled over the 14 to 30 ms span, 190 of 192 classified sequence detectors are silent to the reversed sequence, 99.0 percent.¹

The same dispersion bounds the code. Probing identified detectors with retimed stimuli shows the longest compensable interval tracks Δ almost exactly: 14 ms for Δ of 10 ms, 22 ms for Δ of 20, 31 for 30, 48 for 50, and the bound barely moves when the calcium threshold is varied from 1.5 to 4.5, meaning anatomy rather than the reading gate sets it. Real unmyelinated horizontal fibres are argued to produce a 20 to 30 ms delay spread, which puts the coding bound at roughly the period of gamma oscillations. Precision is the price: at Δ of 20 ms, 87 percent of sequence detectors survive 2 percent velocity jitter, 61 percent survive 5 percent, and 43 percent survive 10 percent. The jitter costing half the matched population falls as one over Δ, from 8.9 percent at Δ of 15 ms to 2.9 percent at Δ of 50 ms, while the tolerated absolute error holds near a millisecond: 1.46 ms at the permissive threshold and 0.82 ms at the strict one. The scheme demands a fixed absolute timing tolerance, the order of the coincidence window itself.¹

Where a skeptic should push

The single most load-bearing assumption is the 1 ms coincidence window in the very cells the column prediction concerns. The paper is admirably frank here: the canonical value sits at or below the bottom of every window measured directly, and no measurement exists in awake animals or in layer 4 stellate cells, the compartment the column construction is about. The windows that are measured come from hippocampus, cerebellum, and dentate gyrus, not neocortical layer 4. If the real layer 4 window is 4 ms, the column diameter prediction widens by a factor of four.¹

Second, the scaling test rests on exactly two points. Monkey V1 gives a median horizontal conduction speed of 0.33 m/s from 156 stimulating-recording pairs across three animals, against an ocular dominance column of about 400 micrometres; the relation predicts 330 to 660 micrometres. Mouse barrel cortex gives a layer 4 intralaminar velocity of 142 plus or minus 76 micrometres per millisecond, against barrels of about 200 to 300 micrometres; the relation predicts 142 to 284. The authors themselves write that two points are not yet a regression, and propagating the uncertainty in velocity and in the crossing angle of afferents widens the prediction to 100 to 1000 micrometres. Worse, monkey V1 conduction speed grows approximately linearly with measurement distance, from the quoted 0.33 m/s median at short separations to several metres per second beyond 5 mm, so the velocity value is partly a property of the distance at which it was measured. And the speed spread within mouse layer 4 is more than half the mean, so the single v in the diameter equation is a central tendency over a very broad distribution.¹

Third, everything is simulation. The detector-class transition emerges in a model with random delays, random connectivity, and a hard calcium gate standing in for dendritic NMDA plateaus; dendritic geometry, ion-channel distribution, and subtype-specific inhibition are abstracted away, and the modelled plateau is far shorter than the 100 to 200 ms of the biological one. The framework also needs a redundant pool of candidate branches, and the paper concedes that exact pairwise matching over N inputs would demand on the order of Δ to the N minus one candidate branches, more than tens to hundreds of real branches can supply for large N, partially alleviated by hierarchy and relay neurons. The strongest version of the claim, the millisecond absolute tolerance, does yield a sharp, falsifiable experimental signature: because compensation depends on delay differences, a 5 percent slowing of conduction should retain about 59 percent of matched detectors while a 5 percent speeding retains only about 48 percent, an asymmetry that timing-precision accounts without delay compensation do not predict. That asymmetry is the experiment to run.

What a myelin-free substrate computes by default

The non-obvious consequence for organoid intelligence falls out of the anatomy switch. A cortical organoid has no myelin. Its projections are the unmyelinated regime the paper assigns a wide delay dispersion, on the order of 20 to 30 ms for horizontal fibres. On this framework, that is not a deficit to be engineered away but a computational classification: wide Δ means the substrate's native detector class is sequence detection, not event detection. Order-selective units should arise spontaneously in randomly wired organoid tissue, provided the tissue can hold conduction delays stable to about a millisecond, and event-type coincidence readouts are what should be rare. This gives organoid training protocols a concrete design rule that no one to my knowledge has stated this way: to make an organoid detect orders and sequences, space your teaching events at an interval near the tissue's intrinsic delay dispersion; to make it integrate synchronous events, you would need to shrink Δ, which in a myelin-free substrate means shorter paths, not faster ones.¹

The millisecond absolute tolerance is the sharpest threat. It is a fixed timing budget, not a percentage, which means it does not relax as the system scales up. Living tissue fails this budget in ways silicon does not: conduction velocity is temperature-, state-, and activity-dependent, and the paper's scaling analysis shows the failure is asymmetric. Slowing conduction dilates delay differences and largely preserves the code; speeding compresses them toward the rounding step and collapses distinguishable patterns onto one. An organoid held half a degree warm, or bathed in a drug that speeds axonal conduction, should degrade its sequence computation faster than the same tissue cooled. That asymmetry is a cheap, decisive functional test of whether a given organoid's readout actually uses delay compensation, and it converts a vague worry about biological drift into a quantitative prediction: protect the substrate from speeding, and it tolerates modest slowing.¹

There is also an underappreciated per-tissue calibration problem. The measured spread of intralaminar conduction velocity, 142 plus or minus 76 micrometres per millisecond in mouse layer 4, is more than half the mean within one layer of one area. If organoid tissue is comparably heterogeneous, every organoid ships with its own private delay signature, its own crossover interval, and its own coding bound near its own gamma period. A stimulation protocol or decoder calibrated on one culture will transfer badly to the next. The paper's learning-by-selection mechanism, in which development and plasticity pick the rare branches whose delays happen to match, suggests the fix: training should select from the redundant pool of pre-existing pathways rather than attempt to impose arbitrary delays. The opportunity is that selection is exactly what closed-loop stimulation can do; the threat is that a substrate whose useful delays are inherited accidents of wiring will batch worse than any silicon part, and its spec sheet will have to be measured, not assumed.

The bottom line

Established in simulation: delay-set dispersion controls the event-versus-sequence detector class, the crossover tracks the coded interval with slope 0.92, and order selectivity is real at 99 percent. Anchored to measurement: coincidence windows of 1 to 2 ms exist in several systems, and two cortical areas fit the velocity-times-window column scaling. Hypothesis: that layer 4 stellate cells actually hold a millisecond window, and that the same scaling holds across more areas. For organoid intelligence the durable idea is that temporal coding class is an anatomical property, and a myelin-free substrate is a sequence detector by default with a fixed, asymmetric, roughly millisecond timing budget. What would confirm the framework: the slowing-versus-speeding asymmetry measured in vivo or in vitro, and more areas added to the column scaling. What would break it: awake-animal measurements of layer 4 integration windows well above 2 ms, or evidence that dendritic coincidence detection in cortex is soft-gated rather than hard-thresholded, since the silent tuned subpopulation that makes the scheme selective exists only under a hard gate.