The Tissue's Shape Did the Computing

We took a 100-micrometre cube of mouse visual cortex — 1,687 neurons and 107,139 synaptic connections, reconstructed by electron microscopy — and passed language over its measured shape as a physical field. No neurons were simulated. Nothing was trained. There are no learned parameters anywhere in the instrument. The tissue discriminated structured language from scrambled language, composed encoded facts exactly, identified relation chains it had never seen, and read which thing acted on which.

The hypothesis is simple to state and was, until now, untested: that comprehension is topological — that it arises from the geometry neural tissue forms, not solely from the activity of the units composing it. If that is true in any measurable degree, then the measured topology of real cortex, with its neurons switched off entirely, should still be able to register and combine linguistic structure presented as a field.

The instrument

A measured connectome becomes a directed flag complex. Its first homology group gives a basis of cycles — closed loops through real neurons, which we call cavities. Those cycles are embedded back in the tissue’s own three-dimensional coordinates, and a text-encoded wave field passes over the tissue — not through it, with no input injected at any node. Each cavity links flux from that field; a discrete Faraday’s law induces circulation around it; and the circulation is read out.

In this sample the complex yields 40 cavities from 1,687 somata, with Betti numbers [13, 40] confirmed independently. Everything the instrument knows about the brain is 40 positions, 40 orientations, and 40 areas, measured from that tissue. There is nothing else to fit.

What was measured

It discriminates — and null substrates do not

The real substrate separates linguistically structured input from scrambled input with an L2 separation of 11.723. Against ten degree-preserving null substrates — same neuron count, same edge count, same degree distribution, rewired — zero of ten came close.

We checked the obvious confound: cavity count correlates with separation at r = 0.81. So we found a null with exactly 40 cavities and compared directly. It loses on both measures. The graph statistics that connectomics papers usually report are not where the function is. The cycle structure and its geometric embedding are.

Facts compose exactly

Reading “A inhibits B.” followed by “B produces C.” as two separate streams, with a small declared state carried between them, lands the tissue in exactly the state produced by reading the combined stream.

Against controls, centred on the comparison-set mean: composition +1.000000000000, wrong verbs +0.925834, identity relations +0.900987, and a scramble at −0.969272. The scramble does not merely score badly. It lands at the opposite end.

It identifies chains it has never seen

Twenty composed queries, none of them in the reference library, classified against 361 candidate relation chains: 20 of 20 at rank 1, margins 0.041 to 0.246. Getting there took ten encoder versions, each forced by a measured failure of the last, and two retractions filed the same day they were caught.

It reads who did what to whom

Twenty forward chains and their twenty argument-swapped counterparts, classified against 722 order-tagged references under two independent readouts:

All 80 rows named the correct verb pair; every miss is an order tag on a thin margin, and the two readouts miss different queries. The declared bar was 40 of 40. It was not met, and under the order’s no-adjustment clause the verdict stands as filed.

The control is the part that matters. A scrambled bag of the same words, which an order-blind readout had happily accepted at rank 1, now fails — its true chain falls to rank 11–17, and all five scramble seeds file under the wrong order tag. Destroying word order destroys identification. That is the direct demonstration that order itself is being read.

What the tissue demanded

The finding underneath the results was not designed in. It was extracted by failure. Each encoder channel proved load-bearing by what broke when it was absent:

What makes it a different kind of machine

There are no parameters. Not parameters sampled without randomness — none at all. The only numbers in the system are 40 cavity positions, 40 orientations and 40 areas, measured off a mouse. Content words derive their identity from a fixed hash of the word itself, computed at encode time, so the vocabulary is open by construction: every possible token already has an identity, including ones nobody has written down.

Which means every answer is traceable. A result can be followed back through the circulation, the flux, the field, to the geometry of a specific piece of cortex. There is no point in the chain where the explanation becomes “the weights say so.” The system is deterministic and reproducible — the same input returns identical values to twelve decimal places across runs — and, more unusually, it is auditable.

The method is part of the result

Every run sat behind a pass/fail bar declared before execution. Every configuration was frozen and hashed, with the hash verified before and after every stage; the digest chain now runs eighteen levels deep and each level still reproduces its predecessor. Every version change was a written amendment against a frozen witness copy, with bit-identity proven over every standing corpus.

Five false or broken results were filed and executed during the first day of work, and each was caught by machinery declared before the result existed. Two identification claims were retracted the same day they were filed — one when the implementing model raised the objection against its own result, one when an anti-circularity census caught floating-point twins. Two runs stopped at their own gates and were reported rather than repaired. A control that behaved impossibly — a scramble outranking all structure — voided an entire run.

What is not claimed

Where this goes

If what computes is cycle structure with a geometric embedding, then nothing requires the cycles be made of mouse. Harvesting a connectome was the fastest route to a topology with the right properties, not a necessary condition. The question becomes which topologies compute well — an engineering search rather than a biological one.

And the timescale is a simulation unit, not a physical constraint. The computation depends on ratios, not on absolute frequency. Every quantity here was measured at a carrier of 1 Hz. Run the same physics at gigahertz and the identical computation happens a billion times faster — which is the case for putting it on silicon rather than in a simulator, where the expensive part is not the physics but the measuring of it.

A provisional patent covering computation using the cycle topology of physical networks as a passive computational medium for structured field-encoded data was filed with the USPTO on 24 August 2026 (application 64/139,859).

Statement

An instrument built from the measured topology of mouse cortical tissue — with no simulated neurons, no training, and no learned parameters — discriminated linguistic structure from noise against ten null substrates, proved its handover state complete to machine precision, composed linguistically encoded facts with a drive and spectrum indistinguishable from single-pass reading, identified the exact relation content of twenty novel composed queries at twenty of twenty, and, once the object was given a physical channel, read who did what to whom at thirty-eight of forty under each of two independent readouts with zero ties — while the scrambled control, which an order-blind readout had accepted, now failed.

Perfect identification required no tissue beyond the original 40 cavities. The capacity was in their frequency structure all along.

The tissue’s shape did the computing. It reads what happened, and — thirty-eight times in forty — who did it to whom.