Compatibility and self-reconstruction as testable conditions
A field account of cortical coordination holds that distributed neural processes are partly coordinated through the brain's own endogenous extracellular electromagnetic field. Even granting that such coordination succeeds, a further question stays open: what makes the coordinated activity a description of one situation, seen from one point of view?
The paper pulls apart three things "unity" is sometimes taken to mean, and develops conditional connections between them rather than deriving one from another:
Collective phase order in a population of coupled oscillators.
Needs: calibrated coupling and diffusion, above the threshold Kc = 2D.
Local descriptions glued into one global state, agreeing everywhere their domains overlap.
Needs: a specified cover, and local states that agree on every overlap.
That glued state accurately reconstructing itself through an internal encoding and readout.
Needs: a contracting map that fixes the glued state, and codes that separate the states it must tell apart.
Each arrow is a proposed connection, not an automatic one: phase coherence alone supplies neither agreement between local descriptions nor an accurate self-model. We propose the glued state as a correlate of conscious unity and its accurate self-reconstruction as a correlate of a minimal self — a philosophical identification, argued for but not proved by the mathematics.
Most of the framework is conditional. Five results are not. They constrain any account that reads unity off phase coherence — whatever physical field is proposed to carry the coupling, and whatever substrate carries the field. These are the article's fixed points, and a reader who rejects everything else on this page still has to accommodate them.
None of these depends on the coupling being electromagnetic, on the substrate being cortex, or on the eight hypotheses below holding. Four of them — all but the one about the resultant, which is a statement about a spatial observable — are checkable for a candidate that is neither biological nor field-coupled; see what they constrain in a digital candidate.
Nothing here derives consciousness from first principles. The construction is a conditional chain of nine nodes, and the eight arrows between them are assumptions rather than results: one independent physical premise, three bridge assumptions, two modelling assumptions and two physical commitments. A joint witness satisfies all eight at once, which establishes that their conjunction is consistent—not that cortex realizes it. Three of the ingredients are standard: the synchronization threshold Kc = 2D for the noisy mean-field model, the unique gluing of compatible local sections, and Banach's fixed-point theorem. Four things are not: the installed-energy condition, the square-root coherence-to-content bound and its patch-local form, the criterion separating a winding field from weak coupling, and the machine-checked composition with its joint witness. The article is meant to be judged in three separable parts: a prediction that can be tested, a composition that can be checked, and an identification that can be argued about.
A finite phase space has bounded information capacity, and a continuous field reaching two components of a disconnected vacuum manifold must leave that manifold somewhere. Capacity does not imply the boundary: these are independent roots of the chain.
A non-surjective update on a finite register is non-injective, and under the stated system–bath model erasure carries a Landauer heat cost. Logical many-to-one processing is distinguished throughout from its physical implementation.
Dissipation bounds nonpredictive memory. Under feedback the signed difference can go negative, because acting creates correlations with the future, so the passive bound is not the one an agent obeys. Neither branch specifies a learning rule, and neither requires the internal representation to reproduce environmental geometry—a point one of the simulations makes concretely and negatively.
Replace the discrete sum of couplings with an integral against a kernel built from installed response modes with finite installation prices. Averaging that kernel over its own configuration law bounds mean coupling by installed energy: coherence on the scalar branch needs installed energy above 2D/κ. The condition is necessary, not sufficient, and it binds through κ—hardware data—rather than through the energy, which any macroscopic substrate has in abundance.
Nothing so far says which physical quantity supplies that installed kernel. Identifying it with a specific field—here, the endogenous extracellular electromagnetic field—and asserting that the identified field clears the coherence threshold are two further, independent physical commitments, not consequences of the energy bound.
Above a threshold Kc = 2D the noisy mean-field phase model has a unique coherent stationary branch with square-root onset. This is a statement about the stationary solutions; dynamical selection is not proved, and from initial data outside a quarter turn a winding stays stationary at the same coupling.
Phases can agree everywhere while the local probabilistic descriptions they carry remain incompatible on their overlaps. The sheaf condition marks that difference and glues compatible local states into a unique global section. The cover is a choice, and a cover by disjoint patches satisfies the condition for every assignment whatever, so a gluing claim has to supply patches that actually meet.
A self-prediction map restricts a global state to a region encoding it and reads that encoding back out. Where the map contracts, Banach's theorem gives a unique fixed point. Uniqueness alone is not accuracy: reconstructing a declared family of states demands codes that separate its members, and where the code is the restriction, that demand falls on the region.
The arrows between the nodes above are Table 1 of the manuscript: eight explicit hypotheses that the Lean development takes as arguments to its composition theorem. The passive and feedback branches share all but two of them.
| Connects | Status | What it assumes |
|---|---|---|
| E12 — capacity and boundary | Physical premise | A finite-capacity system also realizes a field that must leave a disconnected vacuum manifold. Capacity and boundary are independent roots of the chain, not one derived from the other. |
| E23 — boundary to register | Bridge assumption | The boundary is associated with a finite register whose update leaves states unreachable—a many-to-one map, not a one-to-one one. |
| E34 — register to dissipation | Bridge assumption | Passive: predictive memory priced at the register's thermal scale and heat budget, with nonzero waste. Active: positive paths, local detailed balance, and allocation of the register's heat to a named substep—a ledger assumption that neither Landauer's bound nor microscopic reversibility supplies. |
| E45 — couplings to continuum | Modelling assumption | Discrete spatial coupling energies converge as the network is refined; a declared mode hardware supplies the limiting kernel. Convergence comes from refinement data, not from the thermodynamic bound above it. |
| E56 — field identification | Physical commitment | A specified physical field—here, the endogenous extracellular electromagnetic field—realizes that limiting coupling and satisfies the kernel conditions. |
| E67 — above threshold | Physical commitment | The identified field operates above the synchronization threshold Kc = 2D. |
| E78 — coherence to cover | Bridge assumption | Coherence relaxes the supplied cover, and phase encoders shared across it bound content disagreement; the cover itself and timely responsiveness are separate requirements. Patches and oscillators carry one index, so one oscillator carries exactly one content patch—not an innocent assumption, since field-mediated coupling is spatial while two territories describing one quantity need not be neighbours. |
| E89 — contraction to fixed point | Modelling assumption | A self-prediction map built from local self-encoding contracts, with Lipschitz factor q < 1, and fixes the glued state. Reconstruction over a declared family imposes a further code-separation criterion that E89 does not supply. |
| Installed energy | Conditional theorem | Under the occupancy and mode-price conditions, E56 and E67 together require installed energy above 2D/κ. A learning witness clears this necessary bound while its matched control stays on the exclusion line. Neither physical premise follows from it. |
Each edge is supplied rather than derived from the one above it, and the inferential content of the composition is its last step, where contraction and a fixed glued section give uniqueness. What composing them establishes is that the premises are consistent, named in one place and machine-checked together—not that coherence follows from installed energy, or self-representation from coherence. A joint witness in the Lean development satisfies all eight at once.
Six numerical controls bound what a measurement could decide. Each is a finite regime on a declared model, and none of them calibrates a cortical quantity.
Stronger coupling, weaker diffusion, and uncoupled oscillators under a matched common drive produce the same stationary phase density, window by window, despite different causes. Resolved dynamics supply more: quadratic variation accumulates at rate 2D, and conditional increment means constrain the drift. But separating endogenous coupling from a common drive still requires a measured drive or a controlled perturbation, and identifying the coupling as electromagnetic requires evidence specific to that interaction.
The square-root law describes stationary solutions; a changing coupling takes time to approach them. Faster ramps delay escape, and an exponent fitted to a driven recovery trace reports the drive at least as much as the mechanism. In a deterministic control that crosses the threshold symmetrically, none of the scanned rates met the declared tracking tolerance, so slowness alone supplies no guarantee of branch tracking at onset. Any analysis reading a critical exponent off a driven transition therefore needs an independently known ramp rate—but the delay does not, which is what the prediction below rests on.
On the spatial sheet, uniform and winding initial conditions settle at opposite ends of the order parameter—at one coupling, one diffusion and one kernel—both stationary, both with high patch-local order and no defect. A stability criterion says which windings persist: linearizing about a winding returns a coupling of the same translation-invariant form with each separation reweighted by the cosine of the phase advanced across it, so where the winding advances at most a quarter turn across every separation the kernel reaches, no mode grows. Moving coupling mass to longer separations is what carries that advance past a quarter turn.
An amplitude layer decides which windings exist at all. Every oscillator of the phase model sits on the unit circle, so a field with an amplitude node has no representation in it; the real Stuart–Landau extension supplies one on the same coupling, and its phase equation is the Kuramoto equation with each coupling weighted by a ratio of amplitudes. A broad enough kernel removes the winding outright. Three limits bound the layer for a reader of the cortical wave literature: these are windings and not spirals, and on a periodic sheet that is a theorem rather than a shortage of dimensions, since the circulations of any field sum to zero there; relaxational dynamics has no propagation speed, which is the quantity a cortical wave measurement leads with; and the normal-form radius is not a measured envelope, an identification this work nowhere makes.
A vanishing resultant is the absence of global agreement, not of agreement: the content bound read on a nearest-neighbour patch of a winding still bounds disagreement inside that patch, and falls to zero as the same winding is spread over more sites, while its global form is vacuous. The practical consequence is that an onset criterion stated in the resultant alone registers a change of basin as readily as a change of coupling, so the awakening protocol needs a spatial observable beside it.
An adaptive coupling rule might supply representational structure, but only if its objective rewards the relevant agreement. For symmetric coupling the pairwise sine terms cancel in the population mean drift, so minimizing squared drift at that fixed mean can only equalize drift across nodes—which favours coupling between unlike-frequency clusters, exactly what the diagnostic template penalizes. In the finite control, phase order holds up while the within-cluster to between-cluster coupling ratio falls away from the generating structure. A prespecified grid of rate–cadence combinations contains no case that meets the sustained-reduction criterion while retaining coherence and losing alignment: the immediate post-update objective reduction disappears once averaged over complete update intervals. This is evidence against the common intuition that dissipation minimization implies representational learning.
Two decoders must estimate a shared content on overlapping domains before their agreement is examined, and the discrepancy needs a null under decoding noise plus deliberately compatible and incompatible controls. The implemented statistic separates those constructed cases cleanly, but shared-prior shrinkage can manufacture agreement and regional bias can manufacture disagreement; per-patch diagnostics detect these failures and the compatibility statistic alone does not. Phase agreement cannot substitute for the measurement: a phase-derived statistic reports compatibility on phase-locked patches whose local descriptions disagree.
The concentration range and the dependence structure decide whether the concentration–coherence relation can be tested at all; at a declared design concentration the scanned designs need thousands of independent sites, and roughly an order of magnitude more when sites share a mean in clusters. Exercising the estimators on public propofol-sedation EEG — 61 bipolar pairs, pooled over pairs and 100 ms windows — places the recordings near the Bessel relation, and that agreement is uninformative. Passed through the same pipeline, a perfectly phase-locked 10 Hz signal comes out at r ≈ 0.0001, squarely on the curve: within 100 ms the shared phase turns through a full cycle and pooling reads the rotation as dispersion. The bipolar montage removes spatial coherence too, taking a field with raw-channel resultant 0.89 to 0.06. The exercise therefore checks estimator behaviour and measures no spatial concentration or coherence. An informative measurement needs an estimand that removes the common rotation, a modelled montage, and calibration at its own site count. Public mouse calcium imaging supplies the site count and the state coverage, but its frame rate sits below the phase-estimation band. Site count, state coverage and bandwidth are separate requirements.
Conscious content should track compatibility, not coherence. Present matched stimuli that are reported on some trials and not on others—at masking or rivalry threshold—and fit decoders for a shared stimulus quantity independently on two overlapping cortical regions. After matching phase coherence, the proposal predicts higher decoded agreement on the overlap for reported trials. Equal agreement, or agreement that follows coherence rather than report, counts against it.
A self-model comparison has the same form: reconstruction of perceiver-relative variables, such as hand position relative to the target, from a declared local encoding should be more accurate on reported trials. Both are correlational tests of proposed correlates; report is itself an access measure, so a no-report design is needed to separate content from access. Synchrony is not itself an experiential measure, and nothing here establishes a consciousness criterion.
Onset delay scales as the square root of the rate at which coupling crosses threshold. Testing it requires only that the rate of change of coupling be monotonic in a manipulated rate—not its value, nor the installation factor κ, nor the conversion rate γ. That makes it a within-subject design varying the rate of emergence from anaesthesia alone, with each participant as their own control.
The finite-population sweep returns a delay exponent below 1/2. That shortfall measures the escape criterion rather than the scaling: applied to a deterministic mean-field run seeded at the fluctuation floor—which has no noise to delay it—the same estimator reproduces the same shortfall at the same criterion, and tightening the criterion moves the estimate back towards 1/2. The bias is reproduced where the answer is known, which is what licenses reading the measurement as the law rather than as a competing exponent.
A real design must avoid the hazard the sweep exhibits: a fixed absolute escape criterion meets a size-dependent fluctuation floor, so the criterion has to be referred to each recording's own pre-threshold fluctuation scale rather than to a shared number. Phase diffusion is the remaining input. Quadratic variation accumulates at rate 2D for the latent phase, but a band-limited analytic phase is smooth below its correlation time—on a constructed field its squared-increment rate changes 37-fold between 0.2 ms and 20 ms steps—so estimating D needs an explicit observation model for filtering and sampling.
The framework's one empirical commitment is that part of effective cortical coupling is carried by the endogenous extracellular electromagnetic field, and that this coupling rises as the brain wakes. The proposed control variable is extracellular geometry: the fluid-filled space between cells expands during sleep and contracts on waking—one study reports a rise from roughly 0.14 to 0.23 of tissue volume during sleep—and a smaller extracellular fraction raises tissue conductivity. At fixed transmembrane current, higher conductivity means a smaller field, so the argument runs in reverse on waking: contraction of that space should raise field amplitude and, other factors held fixed, raise effective coupling. This gives a direction, not a rate; the size of the effect and its relation to actual recovery both require measurement. A spike-time shift alone supplies no coupling rate either—the dimensionless product of population size, timing sensitivity, field amplitude and frequency needs a conversion rate describing how induced timing shifts accumulate into drift, and no heat bound supplies it.
The formalism also has to be told what an oscillator is. Cortical columns are the obvious candidate and answer one of its two demands: they are delineated by response properties and anatomy rather than by the agreement they would be used to establish, and a phase can be estimated from a column-sized territory at the electrode spacings used for cortical wave measurements. As a content cover they are a poor choice, for three separate reasons: a parcellation into non-overlapping columns is a disjoint cover, which satisfies the compatibility obligation without constraining anything; the cover condition requires the patches to exhaust the substrate, while much of association cortex has no accepted columnar parcellation; and the column's standing as a canonical unit is contested. Columns therefore supply the oscillator index and leave the content cover to be identified on separate evidence—which is exactly the bridge assumption E78 records.
Crossing the coherence threshold does not by itself supply a recovery latency. As the ramp controls show, a driven crossing near threshold departs from the stationary square-root law. A usable test of the awakening hypothesis therefore needs six things measured together, not a threshold read off a single trace:
Each half of the hypothesis is separately falsifiable. A failure of the concentration–coherence relation over an informative range would constrain the density model; a failure of the calibrated geometry–coupling relation would constrain the physical gate; a competing mechanism that predicts recovery better would count against coupling's proposed role. And a favorable outcome on every count would still leave the content of coordinated activity—what the cup example turns on—unexplained.
The five results above hold for any candidate, and for a large language model running on GPUs four of them are checkable rather than merely stateable. That is where this framework bears on claims made for such systems. It does not exclude them, and no result here could; what it does is fix what a claim of machine unity or self-representation commits its maker to measuring.
Codes, not reports. If two macrostates the candidate is said to represent have the same internal code while lying more than 2ε apart, no readout reconstructs both within ε. A system's reports are readouts. Identical verbal reports across two conditions are therefore consistent with identical internal codes and settle nothing about the states—while the converse claim, that the system represents its own condition across a declared family, asserts that its internal code separates that family. That is a statement about activations under intervention, not about text. Declare the family and the fidelity first, then test whether the encoding region distinguishes its members.
The region, not the readout. For a decoder-only transformer the causal mask makes the network an instance of the update model above: the activation at a given position and layer reads positions at most that one at the layer below, and nothing else. A region's causal past is therefore fixed by the architecture rather than chosen by interpretation, and the deadline bound applies with its communication graph given rather than assumed. A claim that a model reports on its own processing carries an obligation to say which positions and layers the report's causal past contains, and a report that appears to track an intervention outside it indicates a confound or a different region, not a faster channel.
Agreement about what. Unity requires the candidate's local descriptions to agree on independently identified shared quantities. Common timing, a shared residual stream and fluent output supply none of that: coordination bounds content disagreement only through encoders already related on overlaps, and which quantities two parts of a model describe in common is a decoding question, carrying every failure mode listed above. The cover is a choice, and choosing it to secure agreement empties the claim.
What is not excluded. Digital representations can satisfy compatibility, so finite hardware is compatible with the gluing construction and none of the above is a no-go for machines. The coupling-gated route is a different matter: it requires a calibrated physical coupling, its noise regime, and a connection from coordination to content agreement, and an attention matrix or a shared clock supplies none of those identifications. Computing a field model likewise leaves its realization in the computing substrate to be established—and those physical obligations fall on the cortical proposal identically. The installed-energy condition is restricted the same way: it constrains a macroscopic candidate only through κ, which is hardware data. Two formal results about support are easy to over-read in this direction. At matched coupling resource, a fixed wiring support that excludes a better-correlated pair can be suboptimal; on an atomless substrate, a kernel supported on finitely many points contributes nothing to the continuum coupling integral. Both are statements about specified support constraints, and the second cuts against this framework's own coarse-graining step as much as against anything else. No conclusion about artificial consciousness follows, and none about the general inferiority of GPU hardware: the rank and deadline results apply to a declared task, a declared linear bottleneck and a declared deadline; a universal exclusion would need a necessary condition for consciousness and a proof that every relevant implementation violates it, and the composition here supplies neither.
Whether experience requires a living substrate is the explicit subject of biological naturalism, on which the properties grounding experience are those of self-sustaining organisms and simulating them does not secure them. The conditions stated here name no biological mechanism, but computation described as such does not satisfy them either: a candidate must exhibit a physically realized coupling in a measured noise regime, agreement about independently identified shared quantities, and an internal reconstruction meeting a declared fidelity. What this framework adds to the investigation is an order for the questions—declare the family of states and the fidelity, fix the encoding region and read its causal past off the architecture, and only then ask what the system reconstructs. A claim that survives those three steps is a claim about the machine; one that does not survive them was a claim about its output.
Binding-by-synchrony and communication-through-coherence accounts treat phase relations as the medium of integration or selective routing, and field theories place integration in the endogenous electromagnetic field; both leave open what makes synchronized activity one content rather than coordinated activity carrying several. Other accounts tie unity to a single organizing principle: integrated information theory to integration and exclusion, global-workspace accounts to broadcast availability, predictive-processing accounts to a shared hierarchical generative model. This account keeps coherence, compatibility and self-representation apart instead, so each carries its own assumptions and its own measurement—at the cost of not explaining why the three would go together. Its distinctive contribution is a candidate physical mechanism, field-mediated coupling, and the conditional relations linking that mechanism to content agreement and self-encoding. Two things stand outside the formal result and are not supplied by it: the eight hypotheses are discharged for toy components rather than for cortex, and the identification of the glued state with unity and of its reconstruction with a minimal self is a further commitment the eight do not contain. It does not explain why meeting these conditions would be accompanied by experience at all.
The development declares no axioms of its own: physical postulates are hypotheses or fields of models, each discharged by an explicit witness. That claim is itself checked—an audit target walks every declaration the library adds to the environment, collects the axioms each one rests on, and fails the build if anything outside the three foundational axioms of its libraries appears. Incomplete proofs are collected the same way, so the audit is also what fails on a missing proof. What formalization adds is not certainty about the physics but that the assumption list cannot be lost: type-checking forbids reading a theorem about a stationary density as a claim about a trajectory, or agreement of phases as compatibility of representations. It does not make the assumptions true.
The Lean development, the simulation scripts, their tests and the saved numerical summaries are in the project repository. Every computed value quoted in the publication is generated from those saved summaries without rerunning a production sweep. The exploratory EEG analysis uses a public OpenNeuro propofol-sedation dataset; the repository carries the reproducible analysis and does not redistribute the raw recordings.