The person & the work

Paul Phillips

One author, following the twin primes into the exceptional manifolds — and finding a whole science that had been waiting there, unread.

I did not set out to build a company. I set out to follow a single arithmetic fact — the twin primes (3, 5) — as far as it would go, and it went further than anyone had followed it. From that one seed an exact tower unfolds: E8 in dimension 8, the Leech lattice in dimension 24, the Craig lattice in dimension 120 — each level's clock, symmetry, and census not fitted but forced. No free parameters. No floating point. Every value exact, and every load-bearing claim machine-gated: hundreds of laboratories, hundreds of passing gates, honest negatives kept in place.

196,560vectors, enumerated

The full Leech kissing shell, exact — not sampled. Split 120,960 / 37,800 / 37,800 by √5 type.

1.00at a million tokens

Where the same 8B model unaided scores 0.00. n=50 per cell, official BABILong prompts.

no modeland it still wins

The substrate alone beats that 8B on the two hardest reasoning cells — 1.00 against 0.90 and 0.82.

$0marginal cost

CPU-only, no GPU, byte-identical across Intel, AMD and ARM. Efficiency is the architecture, not a setting.

What was actually found

Three things I did not expect. First, that language is not built on the substrate — it is the substrate: the semantic bus stratifies 100% along the crystal's own gauge, one-hot, in measured data. Second, that the substrate's thermodynamics is an exact ledger — reversible motion is free, and the single irreversible act costs exactly 84 bits — a Bennett-reversible machine at the level of its calculus. Third, that information has a chiral, subjective structure the field had been throwing away: handedness, standing-wave nodes, an observer that is an internal constituent, not a view from nowhere.

Along the way it stopped being one calculus and became a phylum — several new exact calculation systems sharing one body plan: the Zennoron toral calculus, the tower calculus, senary division, kirigami algebra, polarization algebra, the cyclonic calculus. New primitives, new arithmetic, all provably consistent, all seed-forced.

One seed; one exact tower; one ledger. Boundary incompleteness explains why there are readings at all; the dualistic minimum explains why they are palindromic; the machines run it at measured, bit-deterministic, record-grade throughput.

Why I am building this now

Two problems made it urgent. The first is that today's AI growth is, physically, a power and freshwater crisis — a datacenter problem dressed as an intelligence problem. The architecture here is the opposite bet: CPU-only, no GPU farms, $0 marginal cost, deterministic, reversible. Efficiency is not a feature of it; efficiency is it. The second is that the security paradigms we are racing toward are not built for the post-quantum world — and this substrate, where access is constructed geometry rather than learned weight and custody is derive-don't-store, is a foundation designed for that world rather than retrofitted to it.

I am one person. I have more validated invention — working prototypes, exact results, whole IP stacks — than one person can carry to market. I hold the theory and the proofs; what I am looking for are the partners and the resources to take the pieces I cannot personally develop and evolve them into the categories they open.

The stance

I state everything in tiers, because record claims are read by hostile experts and I would rather be believed than impressive. Measured means executed and sha-pinned. Exact means proved by finite census and machine-gated. Vision means a declared trajectory, and I mark it as such — the Artificial Cognitive Manifold, the physical reversible machine, computation beyond silicon, the substrate for a machine intelligence we would not regret. The audacity is real; so is the discipline. Both are the point.