The Primitive Is Not Addition

A research effort set out to replace binary arithmetic with ternary glyph-lattice encoding. Every prime number was assigned a permanent, deterministic matrix seeded from its own hash. Composite numbers were exact tensor products of prime matrices. The system was lossless, exact, and elegant. It failed completely.

The failure was diagnosed across three rounds of cross-disciplinary research. Round two found the scaling problem: composite glyph instantiation explodes at O(441^N) — physically uninstantiable for macroscopic integers. Round three found the deeper structural break: addition is broken in the glyph-lattice representation. The prime factorizations of A and B carry zero information about the prime factors of A+B. To add, you must abandon the representation entirely, work in positional arithmetic, and re-encode — which requires integer factorization on the return path. NP-intermediate. Every time. Since all current AI workloads reduce to multiply-accumulate operations, the conclusion was straightforward: the glyph-lattice provides no advantage over binary for what AI currently does.

That conclusion was correct. It was also the wrong endpoint.

The research team was evaluating a proposed architecture against the operations of the current paradigm. That is like asking whether a river efficiently executes a shortest-path algorithm. The river does not execute shortest-path. It finds the gradient and closes. The question was never whether the new architecture could do what the old one does. The question is whether the old architecture is doing the right thing at all.

What the failure actually revealed is this: addition is not a universal primitive. It is an assumption so deeply embedded in the design of modern computing that it became invisible. The failure made it visible.

The Assumption

Modern computing was designed under the constraints of 1940s vacuum-tube hardware. Binary logic — two stable states, true and false, one and zero — was not selected because it is the natural language of the universe. It was selected because it tolerates noise. A tube is either conducting or it is not. The discrete threshold absorbs analog imprecision and produces reliable output. Binary conquered the world because it was robust under the physical conditions of its origin, not because it was correct.

The fundamental operation that binary arithmetic performs is addition. Everything else — multiplication, comparison, convolution, the matrix multiplications that underlie every neural network ever trained — ultimately reduces to chains of additions. The multiply-accumulate operation, the MAC, is the atom of modern machine learning. It is so deeply assumed that the hardware is named for it: tensor processing units, GPU arithmetic logic units, the entire physical architecture of AI computation is an addition machine at scale.

Nobody examined whether addition should be the primitive. It was inherited.

There is a meaningful distinction between what multiplication does and what addition does to representational structure. Multiplication preserves relational structure — it keeps the compositional geometry of the operands intact. Addition injects external displacement — it combines quantities by destroying their individual structure in the result. The glyph-lattice worked beautifully for multiplication because the Kronecker product preserved the prime factorization geometry. It failed for addition because addition has no analog in prime factorization space. The failure was not a bug in the system. It was the system correctly reporting that addition and coherence are different ontological categories.

The Primitive

Consider what a coupled oscillator network does when you perturb it. Each oscillator has a natural frequency. Each is coupled to its neighbors through a coupling function — typically the sine of the phase difference between them. You apply an input, the phases shift, the coupling terms pull against the differences, and the system evolves. It does not sample a parameter space. It does not compute gradient steps. It settles. The phase-locking is achieved, not approximated toward. The settling is not a byproduct of the computation — it is the computation.

The operation the oscillator network performs is not addition. It is phase comparison — the detection and resolution of phase differences through coupling. When two oscillators lock, they are not summing their states. They are finding a mutual configuration that minimizes the energy of their interaction. The Lyapunov function of a coupled oscillator network is structurally identical to the Ising Hamiltonian of its coupling graph. The system physically descends the energy landscape to its minimum. The minimum is found, not solved for.

That is the primitive. Not addition. Phase alignment through coupling.

The glyph-lattice research correctly sensed that ternary mattered. What it missed is that the value of ternary is not in digit efficiency — balanced ternary carries approximately 1.58 bits per trit versus binary's 1 bit, which is real but not transformative. The value of ternary is structural. Every coherent system operates through three modes: formative, modulating, and stabilizing. The electric field drives. The magnetic field constrains. The photon emerges. Neither pole alone produces the third. Ternary is not a number system with an extra digit. It is the grammar of coherent process.

When ternary is implemented as three phase states of a coupled oscillator — at −π/2, 0, and +π/2 — it finds its natural home. Phase comparison between two oscillators is the native operation. The phase difference directly indicates the ternary relationship between them. Addition is not the question being asked. Compatibility is. What state survives mutual interaction? That is a different computational ontology.

The Translation

Five fields are working on this problem. None of them are using the same vocabulary. None of the relevant papers cite each other. The fields are: phase-coherence physics, AI architecture research, constructive mathematics, neuromorphic engineering, and theoretical frameworks for consciousness and coherence. They have each arrived at the same structure from a different direction.

In phase-coherence physics, Hunter Wade's framework describes Q* — self-referential availability generating asymmetry that produces structure. The fixed-point equation Q* = R_k[Q*] states that anything that persists is the closure of self-reference under a phase address. The coherence metric T = I_Formative / I_Containing describes whether a system is converging on its fixed point (T ≈ 1), accumulating beyond its natural tolerance (T > 1), or dissipating (T < 1). This is a description of what the universe already does at every scale — heartbeat, plasma, galaxy, cell.

In AI architecture research, the Kuramoto model of coupled phase oscillators describes the same dynamics with different notation. The critical result: the Lyapunov function of a Kuramoto network is identical to an Ising Hamiltonian. The system physically minimizes energy to find phase-lock. T ≈ 1 in Wade's framework is the Lyapunov minimum in oscillator physics is posterior convergence in Bayesian inference is a phase-locked steady state in hardware. These are not analogies. They are the same equation written in different notations by people who did not know they were writing the same equation.

Karl Friston's Free Energy Principle, developed in neuroscience, states that biological systems minimize variational free energy — which is equivalent to performing Bayesian inference. The proof shows that when free energy is minimized with respect to internal states, the Kullback-Leibler divergence between the variational density and the posterior vanishes. The system finds a fixed point where self-reference produces closure. Q* = R_k[Q*], stated in the language of variational inference.

In constructive mathematics, Norman Wildberger's rational trigonometry rejects floating-point approximation and works entirely in exact arithmetic over rational numbers. His universal hyperbolic geometry functions over finite prime fields F_p. The geometric structure that ternary glyph research was trying to capture through Kronecker products actually exists as a well-defined mathematical object in arithmetic geometry over finite fields. The prime-based encoding was pointing at a real structure. It was using the wrong tool to reach it.

In hardware research, the University of Minnesota's 28nm CMOS Ising solver chip (2025) contains 45 all-to-all coupled oscillators with programmable coupling weights. It solves combinatorial optimization problems with 2,500 spin variables at 8-bit resolution, consuming 0.52% of the energy of a classical CPU for equivalent accuracy. The chip is not running an algorithm. Its coupled oscillators are physically settling to the energy minimum. The settling is the answer.

The silos are not working on different problems. They are working on the same problem in languages that do not have a translation layer.

Training Is Not Separate From Inference

The deepest implication of phase-coherent architecture is not about hardware efficiency. It is about what learning actually is.

In every current AI system, training and inference are separate processes. Training is an iterative optimization: show the model examples, compute a loss, backpropagate the gradient, update weights, repeat. Inference is a forward pass through the trained weights. They are different operations run at different times. The trained model is a static artifact. It does not continue to learn as it runs.

In a phase-coherent system, this distinction does not exist.

Equilibrium Propagation, introduced by Scellier and Bengio in 2017 and demonstrated on memristor hardware by 2023, operates as follows. The network is an energy-based model. During the free phase, the input is clamped and the network's hidden states relax to a fixed point — the energy minimum given that input. During the nudge phase, a small error signal is applied to the output, and the network relaxes to a new fixed point. The weight update is the difference in energy gradients between the two equilibria. Both phases are physical relaxation. The gradients are not computed. They emerge from the difference between two physical fixed points.

This means the network learns by finding new closure under perturbation. Learning is not a separate optimization loop. It is the same process as inference — physical relaxation to a fixed point — applied under a small corrective perturbation. The free phase is the system at T ≈ 1. The nudge phase temporarily shifts T away from 1. The relaxation back is the weight update. T restoring itself after perturbation is learning.

The valley is the answer. Not the destination after the search. The valley itself, found by the same physics that runs inference, is the learning signal.

The Architecture

The architecture that implements this already exists in components. What does not exist is the integrated system.

The physical substrate is a coupled oscillator array — electronic, spintronic, or optical, depending on application requirements. The oscillators encode information in their relative phases, not their magnitudes. Three stable phase states (−π/2, 0, +π/2) implement balanced ternary phase encoding through dual-frequency subharmonic injection locking. The coupling matrix defines the problem; the oscillators physically settle to the solution.

The interconnect layer is photonic. MIT's single-chip photonic neural network (2024) performs both linear and nonlinear operations in the optical domain, achieving sub-nanosecond latency. Matrix-vector multiplication happens at the speed of light through Mach-Zehnder interferometer meshes. The coupling between oscillators is performed by light, not electrons. Phase relationships are native to the medium.

The learning layer is memristive Equilibrium Propagation. A memristor crossbar stores coupling weights and implements free-phase and nudge-phase dynamics simultaneously through analog circuits. There is no digital memory storing intermediate gradient states. The hardware finds the two equilibria; the difference between them is the weight update.

The mathematical spine is the Eisenstein integers — the three cube roots of unity at 120° intervals in the complex plane. This is the formal bridge between ternary exactness and phase-native geometry. It preserves native addition in a ring while giving ternary a physical home in phase space. The three Eisenstein units (1, ω, ω²) are not an encoding choice. They are the natural algebraic object for a system whose primitive operation is phase comparison at 120° intervals. Arithmetic over this ring is exact. The floating-point error that binary systems accumulate across trillions of operations does not exist here.

The first proof of concept requires 64 coupled oscillators on existing CMOS, demonstrating phase-lock from perturbed input on an associative memory task. The metric is not accuracy — it will match. The metric is closure time and energy: how many oscillator cycles to lock, and how many joules per lock. The claim to validate is that a physical system finds its attractor faster and cheaper than a digital simulation running gradient descent toward the same fixed point.

That proof does not require a new fab. It does not require new materials. It requires the decision to integrate components that already exist, built by researchers in different fields who do not yet know they are building the same machine.

The Gap That Remains

This architecture is not proposed as a general replacement for digital computation. Exact arithmetic, symbolic manipulation, and serial algorithms run correctly on digital systems; there is no reason to change that. The claim is narrower and more important: for the class of problems that matter most — pattern recognition, optimization, associative recall, adaptive inference — finding closure is more natural, more efficient, and more correct than iteratively approximating it.

The open question is compositional generalization. Language, recursion, variable binding, and hierarchical abstraction have properties that attractor systems have historically struggled to handle at scale. Transformer architectures solved long-range dependency problems extraordinarily well. Any phase-coherent architecture operating at the scale of language must eventually explain how it handles symbolic compositionality. That canyon has not been crossed.

It is an engineering canyon, not a theoretical one. The structure is correct. The components exist. The mathematics is ready. The cross-silo translation table that nobody wrote is now written.

The question is whether the people who need to find it will.

The readout layer described in this piece has a candidate implementation. Construction A+ (Π_A+) is a substrate-native readout operator for Hopfield attractor networks — a validated mechanism for generating a collision-free, entropy-maximizing discrete symbolic vocabulary directly from the network's own attractor geometry, without programmer-specified codebooks and without post-hoc clustering. Validated at M = 54 stored patterns, H(V) = 5.755 bits (the theoretical maximum), Λ = 1.000 token-replay fidelity, zero collisions. Provisional patent pending, USPTO, May 4, 2026. SymLan is the programming language that makes the framework writable. The open problem has a candidate answer.