Measured results and evaluation
The evidence behind the architecture described on the main page: the progress record and certification ladder, the seven demonstrations, the dated self-evaluation with the ranking of 61 developments, the encrypted-in-use research design, and the mathematical essays. “Research scale” marks results obtained on research-scale tasks, before the production model.
Progress to Certification
Every commitment of the programme, marked as it stands. “Complete” means built and passed its pre-registered test; “research scale” means on research-scale tasks, before the production model.
- FoundationsLecture notes, formal verification layer and reproducible test harnessComplete
- FoundationsLanguage-level self-improvement prototypeComplete
- FoundationsGRAIL-native research buildComplete
- Ψ-operator — causal engineCausal identification with machine-checked derivationsComplete · research scale
- Ψ-operator — causal engineCounterfactuals and inverse design with memoryComplete · research scale
- Ψ-operator — causal engineSafe behavioural editing without retrainingComplete · research scale
- Ψ-operator — causal engineCausal reversal and action-order demonstrationsIn progress
- CEAS — global coordinationCoordination scaling under the locality boundComplete · research scale
- CEAS — global coordinationCoordination on a Ramanujan graph (a sparse network with the best possible connectivity for its wiring): averaging across 2,184 units took 77 rounds, against 3,417 on a wrap-around grid of comparable size and the same wiringComplete · research scale
- CEAS — global coordinationLearned global coordination at scaleIn progress
- GRAIL — invariant geometrySymmetry generalisation to unseen group elementsComplete · research scale
- GRAIL — invariant geometryAutomorphic kernel inside the architectureComplete · research scale
- GRAIL — invariant geometryHyperbolic symmetry group checked three ways: element counts at lengths 1–5 match the predicted growth exactly, the growth rate matches the group’s fractal dimension to within 10⁻⁴, and symbolic simplification matched exact matrix arithmetic on 100,000 of 100,000 random pairsComplete · research scale
- GRAIL — invariant geometrySelberg trace formula on the modular surface, a classical identity linking its resonant frequencies to the lengths of its closed loops, checked numerically to within 3 × 10⁻¹⁵Complete · research scale
- GRAIL — invariant geometryRetrofit of existing models (MIA Tiers 0–2)Complete · research scale
- Integration and self-improvementSeed architecture assembled: the three components and five supporting modules (symbolic, spectral, arithmetic, language, self-improvement), linked by 36 cross-checks, each shown to reject a deliberately broken inputComplete · research scale
- Integration and self-improvementFault-injection test of the 36 cross-checks: across all 7,806 combinations of up to three faults, a build-up to all 36, and every on/off pattern of any four (942,480), exactly the faulted checks fired and no healthy one didComplete · research scale
- Integration and self-improvementTriadic integration and the complete ablationComplete · research scale
- Integration and self-improvementClosed self-improvement loop with a machine-checked gateComplete · research scale
- Integration and self-improvementWorld-model self-improvementComplete · research scale
- Integration and self-improvementAll 23 test batteries, including multi-cycle improvement and scaling exponentsComplete · research scale
- Integration and self-improvementSelf-improving control of a drifting physical systemComplete · research scale
- Integration and self-improvementIntegrated-seed evidence (Cert Level 5): all 23 test batteries, all four Pearl rungs, GRAIL orbit generalisation, scaling exponents, self-improving control, hidden benchmark committed in advanceComplete · research scale
- Integration and self-improvementPublic audit bundle for independent replicationIn progress
- ProductPublic deployment with free user accessIn progress
- ProductSelf-improvement applied to production codePlanned
- Long horizonBreadth: cluster-scale evidence across ten or more domainsIn progress · groundwork
- Long horizonPhysics correspondences; quantum and topological substratesResearch
#lang causal-dsl enforcement, memory, audit. Homoiconicity + call/cc for RSI reversion. Chez Scheme performance underneath.Racket + Lean 4
CompleteThe architecture’s own primitives, implemented from first principles
CompleteSWI-Prolog + Lean 4
Complete · research scaleDemo E
Complete · research scaleDemos A + B · public release with the product
In progressDemo C
Complete · research scaleDemo D
Complete · research scaleCert Level 3
Complete · research scaleCert Level 4 · Demo F
Complete · research scaleCausal structure, symmetry and temperature schedule
Complete · research scaleMulti-cycle improvement and scaling exponents
Complete · research scaleCert Level 5
Complete · research scaleEvery script, result and certificate
In progressFastAPI · Oracle Cloud ARM · Supabase
In progressCert Level 6
In progress · groundworkSeven discriminating demonstrations
Each targets an architectural difference, not a benchmark gap.
A dataset where X appears to prevent Y in observed data, but causes Y under intervention. Correlation-only predictors give the wrong direction, and the open-source language model tested never recognised the cases that cannot be identified. The Ψ-causal engine computes P(Y|do(X=x)) by SCM mutilation — correct whenever the effect is identifiable.
“Open drawer then grasp” ≠ “grasp then open drawer.” Order-insensitive models treat these as the same, and the open-source language model tested chose the same order for every pair. The Ψ-causal engine has [Ta, Tb] built in — the commutator is measurable and non-zero.
On N-bit global parity, CEAS reduces coordination rounds via the rank-one collective variable φ(t). Crossover with local attention visible at a measurable scale threshold on a reproducible chart.
Train on a set of group elements; test on unseen elements from the same group. GRAIL: zero error gap. Euclidean transformer: degrades. I(gq, gk) = I(q, k) holds by construction, not by approximation.
Given a target state, compute the action sequence that causes it. Learned inverse models pattern-match forward — they do not invert physical dynamics outside their training distribution. The Ψ-causal engine solves this by construction.
A defined minimum of closed-loop cycles with no human edits inside the loop. Lean 4 gate active throughout. Once the product launches, user queries are designed to feed the continuous fine-tuning cycle.
A controller that keeps its own physics current. As the machine it drives wears and drifts, the closed loop revises its causal model of the dynamics, admits only the terms that predict what happens next and drops the rest, and the Ψ-causal engine plans every action through the revised model. A model fitted once and frozen falls behind; the self-improving loop does not.
A complete graph has no centre: every face touches the other eight, so the face drawn in the middle shows a role, not a position the structure forces. Two placements are defensible.
| Centre | What it shows | A picture of |
|---|---|---|
| Number theory | The common measure: every face is checked against the same zeta and trace identities. | how the system checks itself |
| Language (the edit monad) | The common medium of action: every change to any face is written as a program, checked, then applied. | how the system is designed and steered |
How the nine faces work together
The nine faces are run the way an intelligence community runs its agencies. Two rules pull in opposite directions, and both are kept.
- ◆Compartmentation (avoids entanglement). Each face writes its own report on each candidate from its own data, without seeing the others’ reports. The reports stay independent sources, and no face’s conclusion gets repeated back through another face.
- ◆Fusion (uses every connection). The reports are combined using the methods of the intelligence community:
- ◆Source grading: each face carries a track record of how often its call agreed with the gate’s verdict. Its weight comes from that record, not from an assumption.
- ◆All-source score: the reliability-weighted sum of the faces’ reports, with abstentions carrying no weight.
- ◆Corroboration: a candidate gets a complete trial only when at least two positively graded faces support it.
- ◆Competing hypotheses (Heuer): candidates are also ranked by the weighted evidence against them, since disconfirming evidence is what separates hypotheses.
- ◆Collection management: trial slots left over go to the candidates the faces disagree on most, so the system learns where it is least sure.
- ◆Counter-intelligence: the arithmetic face can veto a candidate whose certificates fail. The trace store can discredit the spectral reports when the spectral record is inconsistent.
- ◆After-action review: the gate’s verdicts update every face’s grade. The fused score’s predictive power is compared with each single face’s, which tests whether fusing helps.
In the two phases these rules govern both directions of the map: which program, read out of a trained model, enters the language, and which program from the language is compiled into a new model.
K9 faces when the model learns nuclear physics
K9 is the network of Logarchéon’s nine coupled components. When the tokens are nucleons, levels or reactions instead of words, the faces take a physical meaning. Complete analysis →
| Face | Meaning for nuclear data |
|---|---|
| Spectral | Model spectra compared with measured level statistics |
| Number theory / Symbolic | Counting shell-model configurations; periodic-orbit sums checked against exact spectra through the trace store (the store counts closed geodesics of the model's group, so the link to a Hamiltonian's orbits is an analogy) |
| CEAS | Inverse temperature of a Gibbs weighting over levels, held near its critical value with a margin; a nuclear temperature when the scores are level energies |
| MIA/GRAIL | Exact Lorentz invariance for collision data; twins as transformations that keep observables and speed the approximate solve (DL ≠ LD) |
| Ψ | Evaluation of reaction and structure models; attribution |
| RSI with the gate | Model refinement accepted only on statistical evidence; a twin jump is one proposed move |
| Arithmetic | Provenance of the data library (hash chain, fingerprints) |
| Language | Programs of moves reused from light to heavy nuclei (the two phases) |
| Langlands | With an arithmetic group, joint Hecke–Laplace eigenfunctions give a kernel basis on the model's own geometry. Whether a fit in that basis helps training is untested; on small models, spectral jumps gave a head start but no lower final loss. The basis does not make training linear: the loss is not linear in the mode coefficients, and the model's parameters live on many copies of the quotient at once. |
In a model of nuclear data, GRAIL’s geometry defines a loss landscape, and its Laplace eigenfunctions are that landscape’s normal modes. When the group is arithmetic, Hecke operators commute with the Laplacian and add labels to those modes; the labels belong to the model’s geometry, not to nuclei. The labelled modes give a kernel basis in which a fit could be made mode by mode. Whether this helps training is untested; on small models, spectral jumps gave a head start but no lower final loss. Each stage →
Not a copy of the brain
The brain is one solution to intelligence, shaped by evolution under the limits of biology: slow signals, unreliable parts, a fixed energy budget, and no means, within biology, of copying itself exactly. A computation is not bound to those limits. The rules a program follows are chosen; mathematics is the one law it cannot break, and physics sets only the price in energy-momentum and spacetime. Every quantity in physics, its constants included, can be expressed in three dimensions: Time (T), Length (L) and Mass (M). The only exceptions are the dimensionless constants, which are pure numbers/ratios, and these ratios may also set limits on a physical AI design. Within those bounds a machine can use what no nervous system can: exact arithmetic, exact symmetry, exact copies, proofs checked by machine. The brain still teaches. Once part of a nervous system is described in mathematics, the description reaches beyond its source: a quantity first recognised as a vector belongs to the family of tensors of every rank, and the mathematics admits all of them at once. In a computer, where the rules are chosen, each of those ranks can be given a meaning in intelligence, either from patterns that appear in data or by a designer’s assignment that is logically and mathematically consistent. The vector is one example; in Logarchéon’s development, one of its approaches treats every quantity of intelligence written in mathematics in this way. A model of the brain is a way into that larger space, not its boundary. The architecture is drawn from that larger space. Its question is not how the brain computes, but how far mathematics allows intelligence to be built.
Where the current build stands
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
Logarchéon’s current build (CEAS–Ψ–GRAIL) placed among 61 AI developments of 2026, each scored on twelve measured quantities in [0, 1], with no judgement scores. Ranking is by Pareto layer and by the mean of the twelve. Status as of 12 October 2026; one development added 31 October 2026. Complete study →
Attention cost at scale (computed)
Replacing dot-product attention scores with GRAIL’s hyperbolic scores changes total training compute by a factor of 1.003–1.14 at GPT-1’s size (117 million parameters), 1.001–1.04 at GPT-3’s (175 billion), and 1.002–1.11 for a dense model of about 400 billion parameters (128 layers, width 16,384, context 8,192). The range depends on how many group words each score sums, from one to five, and on the cost of the distance function. These figures are computed from layer shapes and operation counts, not measured on models of those sizes.
Lorentz-model heads (one hyperbolic distance per head, in ℍdk). Training-compute multiplier against dot-product attention; W is the number of group words summed per score, and c the cost of the distance function per query–key pair, in floating-point-operation equivalents.
| Model | Parameters | Attention share (dot product) | W = 1, c = 10 | W = 1, c = 40 | W = 5, c = 10 | W = 5, c = 40 |
|---|---|---|---|---|---|---|
| GPT-1 | 117M | 3.9% | 1.003 | 1.012 | 1.093 | 1.138 |
| GPT-3 | 175B | 1.4% | 1.001 | 1.002 | 1.030 | 1.038 |
| Dense model (L = 128, d = 16,384, context 8,192) | 412B | 4.0% | 1.002 | 1.006 | 1.088 | 1.111 |
Multi-channel ℍ2 heads (several two-dimensional distances per head), a diagnostic mode run only during troubleshooting; the multiplier applies only while it runs:
| Model | W = 1 | W = 5 |
|---|---|---|
| GPT-1 | 1.23 | 2.24 |
| GPT-3 | 1.08 | 1.44 |
| Dense model (L = 128, d = 16,384, context 8,192) | 1.24 | 2.28 |
On compute cost, the architecture can work at large scale: with Lorentz-model heads, the hyperbolic scores add between 0.1 and 14 percent to training compute at the three sizes computed, and the addition does not grow with model size. The multi-channel diagnostic mode, which costs up to 2.3 times as much while it runs, is used only for troubleshooting.
Frontier-scale models on a single node, with the seed’s safeguards intact (computed)
Running and testing a model of GPT-3 scale or larger usually takes a rack of GPUs. Compressing the model’s stored numbers brings this down to a few GPUs. Compression alone is now widely available; Logarchéon’s contribution is to compress without giving up what the seed architecture adds:
- Exact twins survive compression. The function-identical twin copies used for provenance and protected deployment stay exact after compression. This is proved, not estimated. Ordinary compression breaks it.
- Every compression step is checked and reversible. A step is kept only if the model’s measured accuracy and its causal answers are unchanged, and any step can be undone exactly.
- CEAS-trained models are predicted to be cheaper to store. Holding attention at its critical operating point leaves no idle parameters for compression to remove, and the theory predicts 1.4–2.5 times lower storage cost in the attention’s query and key weights than standard training. This prediction is being tested.
| Model size (parameters) | GPUs needed, no compression | GPUs needed, common 4-bit compression | GPUs needed, best published 2-bit compression | GPUs needed, Logarchéon (estimate) |
|---|---|---|---|---|
| 117 million (GPT-1 scale) | none: a laptop is enough | none | none | none |
| 1.5 billion (GPT-2 scale) | none: a laptop is enough | none: a phone is enough | none: a phone is enough | none: a phone is enough |
| 175 billion (GPT-3 scale) | 5 | 2 | 1 | 1–2 |
| 2 trillion | 58 | 16 | 8–9 | 5–10 |
| 5–10 trillion | 143–286 | 38–76 | 21–40 | 21–59 |
Each GPU is a current data-centre GPU with 80 GB of memory. “GPT-1/2/3 scale” refers only to the published parameter counts of those models, used as reference sizes; no row describes a particular company’s model. Figures are computed, not measured at these sizes, and exclude the memory for long contexts. For the largest models, the GPU count is set mainly by how much information a heavily trained model holds, which no method can compress away.
Next milestones. The seed architecture has been built and tested at research scale; compression is what takes it to large models. Each step is tested in order: models of up to 1.5 billion parameters first, then open models of 14 billion, 70 billion and about 235 billion parameters, the last on a single GPU, with CEAS temperature control, the twin checks and the causal checks applied to each. Results will be posted here as they are measured.
Applied to an open 32-billion-parameter model (measured)
The same controls were applied to an open-weight language model of 32.8 billion parameters. Each change was accepted only when a statistical bound passed on held-out text. Perplexity measures prediction error; lower is better.
| What was done | Result |
|---|---|
| Compressed to 2.6 bits per weight | Held in 11.8 GB of memory; perplexity rises from 10.11 to 12.69 |
| CEAS temperature program, accepted by the gate | Recovers about 5% of the loss from compression (12.69 → 12.55), without changing the stored weights |
| Fact update through a small adapter (0.1% of parameters) | The corrected answer is preferred on 5 of 5 unseen questions (2 of 5 before). No measurable harm on unrelated web text (+0.005 nats per token; 95% bound 0.022) or on 200 general-knowledge questions (69.0% → 71.0%) |
| MIA lift of all 4,096 attention heads to hyperbolic geometry | At curvature 0.1, perplexity changes by less than 0.3% with no retraining, confirmed on web text sealed before the test; at a stronger curvature (0.3) the change is 0.65% on standard text and 1.06% on sealed text |
| Fact placed by one computed change, with no training steps | One low-rank change, computed in a single step and sized using training examples only, placed a new fact better than five training steps did: the margin for the correct answer was higher on all 6 test pairs (+5.1 on average; 95% lower bound +4.3), and the gate accepted it for each of 3 facts set aside before the test |
| Rechecked on web text sealed before the test | On text never used before, the compressed model’s error relative to the original model (1.31) stayed within 5% of the same ratio on standard text (1.26), with the same memory (13.2 GB peak). The hyperbolic lift changed perplexity by at most 1% at moderate curvature; moving every lifted point by a Lorentz isometry left every output unchanged; and the gate refused 3 deliberately harmful changes and accepted 3 harmless ones |
Independent verification · status
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
1 · Measured quantities
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
Twelve quantities, each in [0, 1], computed only from measured numbers; an unmeasured quantity counts as 0.
| System | C | Ψ | G | R | U | H | B | S | V | X | N | D | M12 | M12ᴾ | Seed | Level |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Logarchéon current build | 1 | 1 | 1 | 1 | 0.574 | 0.5 | · | 1 | 0.167 | · | 0.167 | · | 53.4 | 33.6 | 53.4 | 5 (4†) |
| OpenAI math model (unreleased) | 1 | · | · | · | · | 1 | · | · | 0.472 | · | · | 1 | 28.9 | 28.9 | 0 | L1 |
| Axiom Math (AxiomProver) | 1 | · | · | · | · | 1 | · | · | 0.833 | · | · | 0.151 | 24.9 | 24.9 | 0 | L1 |
| AlphaProof | 1 | · | · | · | · | 1 | · | · | 0.833 | · | · | · | 23.6 | 23.6 | 0 | L1 |
| Hyperagents | 1 | · | · | 1 | 0.332 | 0.5 | · | · | · | · | · | · | 23.6 | 18.1 | 0 | L1 |
| Harmonic Aristotle | 1 | · | · | · | · | 1 | · | · | 0.667 | · | · | 0.151 | 23.5 | 23.5 | 0 | L1 |
| ByteDance Seed 2.0 / Prover | 1 | · | · | · | · | 1 | · | · | 0.667 | · | · | · | 22.2 | 22.2 | 0 | L1 |
| Equivariant world model (Wang) | 1 | · | 1 | · | · | · | · | · | · | · | · | · | 16.7 | 8.3 | 0 | L2 |
| Darwin Gödel Machine | 1 | · | · | 1 | · | · | · | · | · | · | · | · | 16.7 | 8.3 | 0 | L1 |
| Anthropic Opus 5/5.5, Mythos, Co-Scientist, GLM-5.3, GPT-6 Astra, Grok 4.6, Kimi K3, Qwen3.8-Max, Sakana AI Scientist | 1 | · | · | · | · | 1 | · | · | · | · | · | · | 16.7 | 16.7 | 0 | — |
| HCP-DCNet | 1 | · | · | 0.95 | · | · | · | · | · | · | · | · | 16.2 | 8.3 | 0 | L1 |
| AlphaEvolve | 1 | 0.238 | · | · | · | · | · | · | · | · | · | 0.521 | 14.7 | 12.7 | 0 | L1 |
| Equivariant sci-ML | 0.667 | · | 1 | · | · | · | · | · | · | · | · | · | 13.9 | 5.6 | 0 | L2 |
| DeepSeek V4-Pro, Huxley-Gödel Machine | 1 | · | · | · | · | 0.5 | · | · | · | · | · | · | 12.5 | 12.5 | 0 | — |
| Self-improving EFMs | 1 | · | · | · | 0.3 | · | · | · | · | · | · | · | 10.8 | 8.3 | 0 | L1 |
| Physical Intelligence π0.5/π*0.6 | 1 | · | · | · | 0.25 | · | · | · | · | · | · | · | 10.4 | 8.3 | 0 | L1 |
| C-JEPA | 1 | 0.22 | · | · | · | · | · | · | · | · | · | · | 10.2 | 8.3 | 0 | L1 |
| H-JEPA (AMI) | 1 | 0.18 | · | · | · | · | · | · | · | · | · | · | 9.8 | 8.3 | 0 | L2 |
| Absolute Zero Reasoner, AgiBot GO-2, AlphaFold 3, Amazon Nova 2, Apertus 1.5, CSML, CausalVAE world-model plug-in, Cohere + Aleph Alpha, DoWhy / y0 / Ananke, ERNIE 5.1, Figure / Skild, Gemini 3.5–3.8 / 4 Argon, Gemini Robotics 2, Genie 3, Hyperbolic nets incl. HELM, Inkling, Intern-S2, Isomorphic Labs, Korea sovereign FMs, Meta Muse Spark, Microsoft MAI, MiniMax M2.7/M3, Mistral AI, NVIDIA Cosmos / GR00T, Nested Learning / Hope (Google), Online causal-rule repair (Bristol), OpenAI research intern, SEA-LION v4, Sarvam-105B, Sber / Yandex, Tencent Hunyuan, UAE K2 Think / Falcon H1R, Unitree, V-JEPA 2-AC, Wayve | 1 | · | · | · | · | · | · | · | · | · | · | · | 8.3 | 8.3 | 0 | — |
| Calibration: GPT-4 on CLadder | 1 | 0.117 | · | · | · | · | · | · | · | · | · | · | 9.3 | 8.3 | 0 | — |
| Calibration: Euclidean transformer inside the build | 1 | — | 0.300 | — | — | — | — | — | — | — | — | — | — | — | — | — |
“·” unmeasured (counts as 0); “—” not applicable. † Level 4 applies if U is read against the loop’s own earlier miss, which rises with wear, instead of against the frozen controller. Discovery values: OpenAI math model, 300 of 719 results with Lean formalisations (openai/math, 7 October 2026); AlphaEvolve, improvements on about 20% of more than 50 open problems (DeepMind, May 2025); Aristotle, Erdős problem #728 with a Lean proof, jointly with GPT-5.2 Pro (arXiv 2601.07421); AxiomProver, Fel’s conjecture with a public Lean proof (Axiom, February 2026).
2 · Gates and the discovery track
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
Levels 1–6 of the certification ladder are unchanged. Three gate qualifiers attach to any level, and a discovery track applies to any system.
Current build: Level 5 (4†) · gate A in progress · gates B and C not started · D0.
3 · The index line
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
M12 · mean of twelve measured quantities
0 20 40 60 80 100
|---------|---------|---------|---------|---------|
^8.3 35 systems
^9.8 H-JEPA
^10.2 C-JEPA
^10.4 Physical Intelligence π0.5/π*0.6
^10.8 Self-improving EFMs
^12.5 DeepSeek V4-Pro, Huxley-Gödel Machine
^13.9 Equivariant sci-ML
^14.7 AlphaEvolve
^16.2 HCP-DCNet
^16.7 10 systems
^22.2 ByteDance Seed 2.0 / Prover
^23.5 Harmonic Aristotle
^23.6 AlphaProof, Hyperagents
^24.9 Axiom Math
^28.9 OpenAI math model
^53.4 Logarchéon current build
M12ᴾ · provenance-weighted
0 20 40 60 80 100
|---------|---------|---------|---------|---------|
^5.6 Equivariant sci-ML
^8.3 42 systems
^12.5 DeepSeek V4-Pro, Huxley-Gödel Machine
^12.7 AlphaEvolve
^16.7 8 systems
^18.1 Hyperagents
^22.2 ByteDance Seed 2.0 / Prover
^23.5 Harmonic Aristotle
^23.6 AlphaProof
^24.9 Axiom Math
^28.9 OpenAI math model
^33.6 Logarchéon current build
4 · The leaders
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
| # | Development | Group | M12 | #ᴾ | M12ᴾ | Seed | Pareto layer |
|---|---|---|---|---|---|---|---|
| 1 | Logarchéon current build (CEAS–Ψ–GRAIL) | Logarchéon | 53.4 | 1 | 33.6 | 53.4 | 1 |
| 2 | OpenAI math model (unreleased) | Science, proof & geometry | 28.9 | 2 | 28.9 | 0 | 1 |
| 3 | Axiom Math (AxiomProver) | Science, proof & geometry | 24.9 | 3 | 24.9 | 0 | 1 |
| 4= | AlphaProof | Science, proof & geometry | 23.6 | 4 | 23.6 | 0 | 2 |
| 4= | Hyperagents | Self-improvement loops | 23.6 | 7 | 18.1 | 0 | 2 |
| 6 | Harmonic Aristotle | Science, proof & geometry | 23.5 | 5 | 23.5 | 0 | 2 |
| 7 | ByteDance Seed 2.0 / Prover | Chinese models | 22.2 | 6 | 22.2 | 0 | 3 |
| 8= | Equivariant world model (Wang) | World models & robotics | 16.7 | 19= | 8.3 | 0 | 2 |
| 8= | Darwin Gödel Machine | Self-improvement loops | 16.7 | 19= | 8.3 | 0 | 3 |
| 8= | Anthropic Opus 5/5.5, Mythos | US frontier models | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Co-Scientist | Self-improvement loops | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | GLM-5.3 | Chinese models | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | GPT-6 Astra | US frontier models | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Grok 4.6 | US frontier models | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Kimi K3 | Chinese models | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Qwen3.8-Max | Chinese models | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Sakana AI Scientist | Self-improvement loops | 16.7 | 8= | 16.7 | 0 | 4 |
Sorted by M12; tied scores share a rank. #ᴾ is the rank on the provenance-weighted reading. Pareto layer 1 holds the 4 developments that no other development matches or exceeds on all twelve quantities while exceeding on at least one. Seed is the third reading: M12 scaled by the pillar core, nonzero only where coordination, causation and invariance are all measured.
5 · Where the main developers stand
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
| Developer | Developments, M12 and rank of 61 |
|---|---|
| OpenAI | OpenAI math model (unreleased) 28.9 (#2); GPT-6 Astra 16.7 (#8=); OpenAI research intern 8.3 (#27=) |
| Google DeepMind and Google Research | AlphaProof 23.6 (#4=); AlphaEvolve 14.7 (#19); Gemini 3.5–3.8 / 4 Argon 8.3 (#27=); Gemini Robotics 2 8.3 (#27=); Genie 3 8.3 (#27=); AlphaFold 3 8.3 (#27=); Nested Learning / Hope 8.3 (#27=) |
| Anthropic | Anthropic Opus 5/5.5, Mythos 16.7 (#8=) |
| xAI | Grok 4.6 16.7 (#8=) |
| Meta and AMI Labs | H-JEPA (AMI) 9.8 (#26); V-JEPA 2-AC 8.3 (#27=); Meta Muse Spark 8.3 (#27=) |
| Microsoft, Amazon, NVIDIA | Microsoft MAI 8.3 (#27=); Amazon Nova 2 8.3 (#27=); NVIDIA Cosmos / GR00T 8.3 (#27=) |
| Harmonic | Harmonic Aristotle 23.5 (#6) |
| Axiom Math | Axiom Math (AxiomProver) 24.9 (#3) |
| ByteDance | ByteDance Seed 2.0 / Prover 22.2 (#7) |
| Moonshot, Alibaba, Zhipu | Kimi K3 16.7 (#8=); Qwen3.8-Max 16.7 (#8=); GLM-5.3 16.7 (#8=) |
| DeepSeek | DeepSeek V4-Pro 12.5 (#21=) |
| Shanghai AI Lab, MiniMax, Baidu, Tencent | Intern-S2 8.3 (#27=); MiniMax M2.7/M3 8.3 (#27=); ERNIE 5.1 8.3 (#27=); Tencent Hunyuan 8.3 (#27=) |
| Chinese robotics | AgiBot GO-2 8.3 (#27=); Unitree 8.3 (#27=) |
6 · All 61 developments
This is a landing page. It states results and criteria only; how each result was obtained, and why the architecture takes its form, are withheld here. Proofs and implementations are rechecked against the stated claims by separate verification passes, run by AI models and agents from different companies, to detect drift.
Show all 61 developments
| # | Development | Group | Level | V | X | N | D | M12 | #ᴾ | M12ᴾ | Seed | Pareto layer |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Logarchéon current build (CEAS–Ψ–GRAIL) | Logarchéon | L5 (4†) | 0.167 | · | 0.167 | · | 53.4 | 1 | 33.6 | 53.4 | 1 |
| 2 | OpenAI math model (unreleased) | Science, proof & geometry | L1 | 0.472 | · | · | 1 | 28.9 | 2 | 28.9 | 0 | 1 |
| 3 | Axiom Math (AxiomProver) | Science, proof & geometry | L1 | 0.833 | · | · | 0.151 | 24.9 | 3 | 24.9 | 0 | 1 |
| 4= | AlphaProof | Science, proof & geometry | L1 | 0.833 | · | · | · | 23.6 | 4 | 23.6 | 0 | 2 |
| 4= | Hyperagents | Self-improvement loops | L1 | · | · | · | · | 23.6 | 7 | 18.1 | 0 | 2 |
| 6 | Harmonic Aristotle | Science, proof & geometry | L1 | 0.667 | · | · | 0.151 | 23.5 | 5 | 23.5 | 0 | 2 |
| 7 | ByteDance Seed 2.0 / Prover | Chinese models | L1 | 0.667 | · | · | · | 22.2 | 6 | 22.2 | 0 | 3 |
| 8= | Equivariant world model (Wang) | World models & robotics | L2 | · | · | · | · | 16.7 | 19= | 8.3 | 0 | 2 |
| 8= | Darwin Gödel Machine | Self-improvement loops | L1 | · | · | · | · | 16.7 | 19= | 8.3 | 0 | 3 |
| 8= | Anthropic Opus 5/5.5, Mythos | US frontier models | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Co-Scientist | Self-improvement loops | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | GLM-5.3 | Chinese models | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | GPT-6 Astra | US frontier models | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Grok 4.6 | US frontier models | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Kimi K3 | Chinese models | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Qwen3.8-Max | Chinese models | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 8= | Sakana AI Scientist | Self-improvement loops | L1 | · | · | · | · | 16.7 | 8= | 16.7 | 0 | 4 |
| 18 | HCP-DCNet | Self-improvement loops | L1 | · | · | · | · | 16.2 | 19= | 8.3 | 0 | 4 |
| 19 | AlphaEvolve | Self-improvement loops | L1 | · | · | · | 0.521 | 14.7 | 16 | 12.7 | 0 | 1 |
| 20 | Equivariant sci-ML | Science, proof & geometry | L2 | · | · | · | · | 13.9 | 60 | 5.6 | 0 | 3 |
| 21= | DeepSeek V4-Pro | Chinese models | L1 | · | · | · | · | 12.5 | 17= | 12.5 | 0 | 5 |
| 21= | Huxley-Gödel Machine | Self-improvement loops | L1 | · | · | · | · | 12.5 | 17= | 12.5 | 0 | 5 |
| 23 | Self-improving EFMs | World models & robotics | L1 | · | · | · | · | 10.8 | 19= | 8.3 | 0 | 3 |
| 24 | Physical Intelligence π0.5/π*0.6 | World models & robotics | L1 | · | · | · | · | 10.4 | 19= | 8.3 | 0 | 4 |
| 25 | C-JEPA | World models & robotics | L1 | · | · | · | · | 10.2 | 19= | 8.3 | 0 | 2 |
| 26 | H-JEPA (AMI) | World models & robotics | L2 | · | · | · | · | 9.8 | 19= | 8.3 | 0 | 3 |
| 27= | Absolute Zero Reasoner | Self-improvement loops | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | AgiBot GO-2 | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | AlphaFold 3 | Science, proof & geometry | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Amazon Nova 2 | US frontier models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Apertus 1.5 | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | CSML | Science, proof & geometry | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | CausalVAE world-model plug-in | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Cohere + Aleph Alpha | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | DoWhy / y0 / Ananke | Science, proof & geometry | L2 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | ERNIE 5.1 | Chinese models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Figure / Skild | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Gemini 3.5–3.8 / 4 Argon | US frontier models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Gemini Robotics 2 | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Genie 3 | World models & robotics | n/a | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Hyperbolic nets incl. HELM | Science, proof & geometry | L2 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Inkling | US frontier models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Intern-S2 | Chinese models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Isomorphic Labs | Science, proof & geometry | L1–L2 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Korea sovereign FMs | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Meta Muse Spark | US frontier models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Microsoft MAI | US frontier models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | MiniMax M2.7/M3 | Chinese models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Mistral AI | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | NVIDIA Cosmos / GR00T | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Nested Learning / Hope (Google) | Self-improvement loops | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Online causal-rule repair (Bristol) | Self-improvement loops | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | OpenAI research intern | Self-improvement loops | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | SEA-LION v4 | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Sarvam-105B | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Sber / Yandex | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Tencent Hunyuan | Chinese models | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | UAE K2 Think / Falcon H1R | Rest of world | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Unitree | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | V-JEPA 2-AC | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
| 27= | Wayve | World models & robotics | L1 | · | · | · | · | 8.3 | 19= | 8.3 | 0 | 6 |
Sorted by M12; tied scores share a rank. Developments absent from the quantified table receive the floor value C = 1.
Generating a better language is formulating a better theory
RSI Level 2 changes the language itself — what can be written and what is rejected before execution.
Each proposed language change passes a gate: it may only make more errors impossible, must keep every known-good program valid, and must improve on the error suite.
The loop cannot regress, always terminates, and blocks over-broad changes.
A research build now runs this loop on a GRAIL-native model. The language grows by abstracting the operations that succeeded, and every extension passes the gate with a certificate checked in Lean 4.
On sixteen tasks the model had never seen, the grown language corrected the model clearly better than the language it started from, without a single gradient step and without loss on the patterns it already knew. The architecture’s own identities held to machine precision throughout, and every edit was an isometry checked by Racket and certified by Lean 4.
When weights and programs live on the same manifold
Weights and inputs share one geometry, so a program edit and a geometric edit can be treated alike. Seven capabilities follow from this; details are in the technical brief.
The GRAIL-native research build realises this. The architecture's own primitives are implemented from first principles, every parameter is a point of the geometry, and the identities the theory predicts — function-identical twin models, and training that commutes with the symmetry group — hold to machine precision. Edits to the model are themselves programs of isometries, checked before they run.
Retaining pre-trained knowledge
A protocol for importing pre-trained knowledge without breaking invariance is specified in the technical brief.
Self-improvement, one module at a time
The other sixty developments ranked above grow or improve in seven ways. Frontier models, from the United States, China and elsewhere, are retrained or post-trained as one whole network. Provers (the OpenAI math model, AxiomProver, AlphaProof, Aristotle, Seed Prover) and the Absolute Zero Reasoner train a single model against a proof checker or a code executor. AlphaEvolve searches over programs scored by an evaluator. Agents (Hyperagents, the Darwin and Huxley–Gödel machines, Co-Scientist, the AI Scientist, the OpenAI research intern) rewrite their own code or research procedure and judge the result on benchmarks. Nested Learning (Hope) updates its own memory at several rates while it runs. Other loops in the ranking repair rules or parameters in place (HCP-DCNet, online causal-rule repair), and self-improving embodied models retrain on the data they collect. World models and scientific models (the JEPA family, Genie, Cosmos, AlphaFold, the equivariant, hyperbolic and causal toolkits) learn by prediction or fitting and do not modify themselves. In their published descriptions, none of these certifies, change by change, that the rest of the system was left as it was.
Logarchéon has developed an original method, distinct from all seven. Every other route stakes the whole system on each change: an update aimed at one skill can quietly damage another, and the damage is found, if at all, only by testing everything again. Logarchéon’s method organises a model into separately upgradable modules, so that each improvement is confined to one module, checked against the skill it targets, recorded, and exactly reversible. In the architecture’s own model class this is proved; for open models it is checked change by change. Improvements therefore build on one another instead of trading off; the growth needed for each new skill is known and costed before any training; and every change leaves the auditable record that high-assurance users need. The method starts from existing open models instead of costly new training, and it rests on machine-checked mathematics, not on benchmark scores alone.
- ◆A complete theory, machine-checked. 193 results on modular, self-upgrading architectures, each verified in Lean 4.
- ◆One skill, one module. In the architecture’s own model class it is proved that a module can be retrained or replaced while every other skill stays exactly as it was, so each change needs to be checked only against the skill it targets.
- ◆Trained models joined without retraining. Two frozen models can be combined by one added control level, whose cost is known in advance and proved to be the least possible.
- ◆A known price for each new skill. How much the architecture must grow to hold each added skill is stated by theorem before any training.
- ◆Stable at any depth. A proved control condition keeps deeper hierarchies of the architecture stable, removing the instability that has limited them.
Models as building blocks. In the architecture’s own model class, trained models combine the way numbers do. A model’s output can be scaled or negated at no added depth. Two models can be added or joined at the cost of one control level, and multiplied at the cost of two. Joining is associative and has a unit, so any number of models give the same result in any grouping. Modules can be added, replaced, removed or moved without retraining the rest. Each operation is machine-checked in Lean 4. Together they form an algebra of models, so a large system can be built up, upgraded or taken apart one module at a time.
Applied to open models (computed). Organising a 32-billion-parameter open model into checked, upgradable modules: about 350 GPU-hours on entry-level hardware, roughly one week on two entry-level GPUs.
Planned service. Skill audits of open models, and certified skill modules for organisations that need one model strengthened on chosen tasks, each delivered with a record that no other skill regressed.
A mathematics of trained models
Logarchéon has developed a new mathematical theory that treats trained neural networks as objects of number theory. Models are given operations, as numbers are: they can be scaled, negated, added, multiplied and joined. They fall into equivalence classes, factor into primes, and carry zeta functions.
The theory runs from numbers to networks without a break. A model with a single unit and weight w is the map z ↦ wz. Under composition, the prime models among those with positive whole-number weights are exactly the prime numbers, so Riemann’s zeta function appears as the simplest case. Larger models, their graphs and their dynamics carry zeta functions of their own. For each of these, the Riemann hypothesis takes a definite form: a theorem in some cases, an open question in others, and in some false.
The theory is axiomatic in the standard sense. Its foundation is the logic of the Lean 4 proof assistant, a standard foundation for mathematics, which uses three axioms: propositional extensionality, the axiom of choice, and soundness of quotients. On that foundation the theory adds no new axioms, only definitions. Models, their operations, their equivalence classes and their zeta functions are defined objects, and every theorem about them follows from those definitions. 193 such theorems are machine-checked. The theory, its operations and its new objects were developed by Logarchéon and are first published here.
- Propositional extensionality (
propext): two propositions that imply each other are equal.axiom propext {a b : Prop} : (a ↔ b) → a = b - The axiom of choice (
Classical.choice): from a proof that a type is nonempty, an element of it can be chosen.axiom Classical.choice {α : Sort u} : Nonempty α → α - Soundness of quotients (
Quot.sound): elements related by a relation are equal in the quotient by that relation.axiom Quot.sound {α : Sort u} {r : α → α → Prop} {a b : α} : r a b → Quot.mk r a = Quot.mk r b
- ◆An algebra of models. Models can be scaled, negated, added, multiplied and joined. Each operation has a proved cost in depth, joining is associative and has a unit, and the outputs that models can hold form a commutative ring.
- ◆Disentangling at linear cost. Measuring how two units of a model interact takes four passes, whatever the number of paths between them. Mapping a large model therefore costs in proportion to its units, not to the number of its paths.
- ◆A complete classification of model equivalences. Every way of identifying two models falls into one of seven classes, and classes acquire primes in only three ways. Twenty-two routes from models to primes and zeta functions are catalogued, with the status of the Riemann hypothesis on each.
- ◆One carrier for zeta functions. The weighted graph of a model carries every Euler product, Riemann’s included. Through five generalisations it also carries the Hasse–Weil, Ihara, Ruelle and Selberg zeta functions. Wherever an operator exists, the Riemann hypothesis becomes a bound on that operator’s spectrum.
- ◆New curves from model dynamics (computed). Feeding a layer its own output defines curves of states that repeat. For a two-dimensional family these curves are irreducible, of genus 3 and 31 at the first two periods, with Galois groups as large as the dynamics allows. A stated conjecture extends this to every period.
New mathematics: theorems and constructions
| Result | Statement | Status | Verification |
|---|---|---|---|
| Operations on models | In the class of polynomial models that return task-indexed families exactly under admissible task codes: if F is returned at depth n1 ≥ 1 and G at depth n2 ≥ 1, with N = max(n1, n2), then cF and −F (c ∈ ℚ) are returned at depth n1; F + G and the join F ∨r G under any routing r at depth N + 1; FG at depth N + 2; and a family returned at depth n is returned at every depth n′ ≥ n. Each result is realised from the two frozen models without changing their parameters. | New | Lean 4 |
| Laws of joining | For any assignment of tasks to three models, the groupings (F ∨ G) ∨ H and F ∨ (G ∨ H) are equal, and routing every task to F returns F. On the machines themselves the laws hold up to relabelling of coordinates. | New | Lean 4 |
| The ring of models | The families returned by some model form a commutative ℚ-subalgebra of ℚ[x]ℕ, closed under every join. Depth is a cost function, not a grading. Complementary routings give zero divisors, so there is no general division. On machines the structure is a bimonoidal (rig) category, whose quotient by equality of outputs is this ring. | New | Lean 4 |
| Path sums and dropout | For layers Mk and node masks mk: (∏k diag(mk) Mk)ij = ∑ over paths p from i to j of ∏k mk(pk) Mk(pk, pk+1). Dropping a unit removes exactly the paths through it, so one pass measures all of them. For two units u, u′: f − f−u − f−u′ + f−uu′ equals the sum over the paths through both, so four passes measure their interaction whatever the number of paths. Single-unit sweeps take L·w passes against wL−1 paths. | New application of a classical identity | Lean 4 |
| Isolation of modules | If the layers are block-diagonal over modules, units in different modules have zero interaction. If modules share a trunk of t layers and their branches above it are closed, units in different branches have zero interaction, and each trunk unit interacts with every branch exactly through the paths that pass through it. | New | Lean 4 |
| Classification of equivalences | Every equivalence relation on models is the kernel of an invariant of (architecture, parameters, history); sorting by the components an invariant depends on gives seven classes. Classes acquire primes in exactly three ways: a product with unique factorisation, closed orbits, or no Euler product at all. An Euler product exists if and only if the class counts by norm equal the counts of multisets of primes. | New | Proved |
| Twenty-two routes to zeta functions | Each route pairs an invariant with a mechanism. Examples: code words under concatenation have the Lyndon code words as primes, ζ = 1/(1 − qu) and prime count (1/n)∑d|n μ(d) qn/d, where the hypothesis holds; lattice codebooks give Epstein’s zeta function, open for ℤ2 and false for lattices of class number greater than one; dropout configurations have zeros on the unit circle under symmetric couplings bounded by one (Lee–Yang); module–weight attachment matchings have real zeros in |x| ≤ 2√(Δ − 1) (Heilmann–Lieb); spanning trees of the module tree are real stable (Choe–Oxley–Sokal–Wagner); commuting edits have generating function 1/I(G, −u) (Cartier–Foata) with real zeros for claw-free conflict graphs (Chudnovsky–Seymour); first-layer activation regions have characteristic polynomials with positive-integer roots when supersolvable (Stanley); closed walks of a shared layer satisfy the hypothesis exactly when they meet the Ramanujan condition, and their zeta function factors over block-diagonal modules, the order of the leading pole being the number of modules at the spectral radius. | New catalogue | Applied |
| One carrier for zeta functions | The cycle zeta function ∏P(1 − w(P))−1 of a weighted graph, over its prime cycles, carries every Euler product: disjoint loops of weight N(x)−s give the Riemann, Dedekind and Serre zeta functions; a bouquet of q loops gives 𝔽q[t]; the functional graph of Frobenius gives Hasse–Weil. Through countable, graded, matrix-valued, path-dependent and continuous generalisations it also carries the Ihara, graph L-, Ruelle and Selberg zeta functions. For a finite graph the hypothesis is equivalent to a bound on the non-leading eigenvalues, which gives a square-root error in prime counts. | New unification | Lean 4 (core); applied |
| Riemann’s zeta inside the framework | One-unit models z ↦ wz compose by multiplying weights. Among positive-integer weights, a model is a composite of two models of weight at least two exactly when its weight is not prime, so the zeta function of this composition monoid is ζ(s); Gaussian weights give the Dedekind zeta function of ℚ(i). | New | Lean 4 |
| Dynatomic curves of a quadratic family | Φc(x, y) = (1 + x2 + 2(1+c)xy − y2, c + (1+c)x2 + (1+c)xy + cy2) is regular wherever (2c+1)2(c2 + 2c + 2) ≠ 0. Its dynatomic curves Y(1), Y(2), Y(3) are irreducible over ℚ, of degrees 4, 12 and 60; Y(1) has genus 3 and Y(2) genus 31, against 0 and 0 for z2 + c. Y(2) branches simply at 12 period-doubling values, at two points over each of 32 saddle-node values, and with index three at c = −1 ± i. | New object | Computed exactly (simple branching confirmed to 110 digits); regularity and point counts over ℤ/3, ℤ/5, ℤ/7 in Lean 4 |
| A Galois orbit at c = 2 | The fibre of Y(2) at c = 2 factors over ℚ as 2 + 10. The 2-cycle {(3/5 + ω, −1/5), (3/5 + ω̄, −1/5)}, with ω2 + ω + 1 = 0, is defined over ℚ(√−3), and Φ2 maps each point to its Galois conjugate. | New | Lean 4 |
| Galois groups (conjecture) | For a generic one-parameter family of regular quadratic self-maps of affine D-space, D ≥ 2, the dynatomic curves Y(n) are irreducible for every n, with Galois group Cn ≀ Sν(n)/n, where ν(n) = ∑d|n μ(n/d) 2Dd. Supported for n ≤ 3: Frobenius factorisation patterns over about 26,000 primes match this group at the sampling-noise floor. | Conjecture | Computed |
| Second eigenvalue of random Markov matrices; golden-ratio threshold of attention operators | Normalise each row of a large random matrix with independent nonnegative entries (mean m, standard deviation σ) so that it sums to one. The resulting random walk forgets its starting point at a rate set by its second eigenvalue λ₂. Proved: √N |λ₂| → σ/m almost surely, for every weight law with a finite fourth moment, including the random walk on a random directed network. This answers a question raised by Bordenave, Caputo and Chafaï in 2008 and settles the conjecture Chafaï stated in 2007 for the Dirichlet Markov ensemble; both had been open since. The lower bound and the circular law for these random walks hold under a finite variance alone, with no density; the circular law had required a bounded density. Under a finite variance alone, √N |λ₂| → σ/m in probability. For a softmax attention operator with independent Gaussian scores at inverse temperature β, it gives |λ₂|·eh/2 → √(eβ²−1)·e−β²/4 almost surely, where h is the entropy rate. The second eigenvalue therefore falls below the entropy scale e−h/2, the upper bound known for random regular digraphs, for β < √(2 ln φ) ≈ 0.981, and rises above it beyond; φ is the golden ratio. Preprint: doi:10.5281/zenodo.23903641 | New | Lean 4 (112 declarations: every deterministic lemma and the deterministic estimates of the proof); the random-matrix inputs are published theorems |
Lean 4: machine-checked. Computed: exact computer algebra, with any numerical step named. Applied: a published theorem applied to a new object.
Ways to form equivalence classes of models
| Class | Looks at | Two models are equivalent if … | Primes? | Primes, and how they arise | Zeta function and its RH | Logarchéon’s contribution |
|---|---|---|---|---|---|---|
| 1 | Architecture | their graphs of units are isomorphic; their module trees (children and parents) are isomorphic; they have the same depth and width profile | Yes | prime architectures under the Cartesian product (Sabidussi–Vizing) | architecture counting series; no known locus | Defined the relation on models; identified the prime factorisation that applies |
| 2 | Parameters | their weights differ by a symmetry (unit permutations, rescalings, isometries of the readout geometry, relabelling) | No | none (group orbits) | none | Identified the symmetry groups of models, including the isometries of the readout geometry |
| their weights agree modulo a prime, or are Galois or Frobenius conjugates | Yes | Frobenius orbits | Hasse–Weil; RH holds (Deligne) | Defined models over finite fields; applied Deligne’s theorem to them | ||
| they quantise to the same code word (weight codebooks of the kind used to compress large models) | Yes | Lyndon code words under concatenation | 1/(1 − qu); RH holds | New route: stacking quantised models concatenates code words, so the Lyndon primes and their exact count apply | ||
| they fall in the same cell of a lattice codebook | No | none | Epstein; open for ℤ2, false for some lattices | Showed that the codebook decides whether a hypothesis exists at all | ||
| their shared layers have the same spectrum | No | none | spectral zeta; no known locus | Defined the relation | ||
| they lie in the same connected set of near-optimal weights | No | none | none | Placed it in the classification | ||
| 3 | History | they lie on the same periodic orbit of training | Yes | periodic orbits of the training map | Ruelle type; with quantised weights, poles on the unit circle | Proved the period-counting identity for quantised loops (Lean 4) |
| their edit sequences agree up to reordering of commuting edits | Yes | Lyndon traces (Cartier–Foata) | 1/I(G, −u); real zeros when the conflict graph is claw-free | Proved that edits on disjoint modules commute (Lean 4); identified the trace monoid and the theorems that apply | ||
| 4 | Architecture and parameters (behaviour) | they give equal outputs on every task, on a task set, or within a tolerance | Yes | irreducible outputs under product; indecomposable models under composition (Ritt) | open | Proved the operations and laws on models (Lean 4); showed unique factorisation fails for families |
| after quantising weights and activations to finite grids (integers; Gaussian or Eisenstein lattices for complex values), they have the same quantised graph: the same connections, quantised weights and activation tables | Yes | connected quantised graphs under disjoint union; on grids that are finite fields (ℤ/p, ℤ[i]/(π)), Frobenius orbits | finite counts at each grid size give a counting zeta function; on finite-field grids, Hasse–Weil with RH (Deligne) | Logarchéon’s quantisation method: it turns uncountably many real-valued models into finitely many classes at each grid size, so classes can be counted and zeta functions defined | ||
| with quantised activations, they compute the same transducer, or are bisimilar | Yes | periodic orbits; Krohn–Rhodes factors under cascades | Artin–Mazur; poles on the unit circle | Identified the route; proved the orbit count (Lean 4) | ||
| they are one-unit models z ↦ wz with the same weight | Yes | rational primes; Gaussian primes | Riemann ζ(s); Dedekind ζ of ℚ(i) | New theorem (Lean 4): composition multiplies weights and the primes are the rational primes, which places Riemann’s ζ inside the theory | ||
| their outputs define the same number field or Galois group | Yes | prime ideals | Dedekind and Artin; generalised RH, open | Defined the relation on outputs | ||
| their activation regions have the same combinatorial type | Yes | irreducible hyperplane arrangements | characteristic polynomial; positive-integer roots when supersolvable or free | Identified the route and its conditions | ||
| the closed walks of a shared layer agree up to rotation | Yes | primitive cycle classes | cycle zeta 1/det(I − W); RH iff Ramanujan | New: the cycle zeta factors over disentangled modules, with leading-pole order equal to the number of modules; rotation and repetition laws proved (Lean 4) | ||
| their interaction graphs (four-pass measurement) are isomorphic | Yes | primitive interaction cycles | Ihara type; RH iff Ramanujan | New: the four-pass interaction theorem (Lean 4) and the route built on it | ||
| their dropout configurations have the same weights | No | none (zero locus only) | Lee–Yang: unit circle under symmetric couplings bounded by 1 | New theorem (Lean 4): for one hidden layer the dropout energy is exactly a lattice gas; trained couplings have mixed signs, so Lee–Yang’s condition fails, as measured | ||
| their tangent-kernel or curvature spectra agree | No | none | spectral zeta; no known locus | Placed it in the classification | ||
| 5 | Architecture and history | they are built by the same joins, up to regrouping | Yes | indecomposable modules under join | prime count; open | Proved that joining is associative and has a unit (Lean 4), and that connected modules factor uniquely |
| 6 | Parameters and history | they differ by edits confined to one module, or carry the same certified-property record | No | none | none | Proved that edits confined to a module leave the rest unchanged (Lean 4) |
| 7 | All three | they are the same model with the same history | No | none | none | Placed it in the classification as the finest relation |
Rows that carry primes are highlighted: 14 of the 22 ways. The list is complete by construction: every equivalence relation is the kernel of an invariant, and every invariant depends on a nonempty subset of {architecture, parameters, history}, which gives exactly seven classes (a classification developed and proved by Logarchéon). Within a class the relations can be refined without end. Classes acquire primes in only three ways: a product with unique factorisation, closed orbits, or not at all.
The Riemann hypothesis for the reductions of these curves modulo primes holds by Weil’s theorem; the open questions are their irreducibility, Galois groups and the growth of their genus.
From theory to implementation
Theory-to-implementation mappings are provided in the technical brief.
Nuclear physics, fusion and gravitation
Logarchéon’s AI designs draw on statistical physics, hyperbolic geometry and causal inference, the same mathematics used in nuclear physics, fusion and gravitation. Some of the links are exact identities; others are analogies that still require testing. Complete analysis →
Where the designs map onto nuclear physics
New physics: theoretical results
| Logarchéon piece | Nuclear counterpart | Status | Verification |
|---|---|---|---|
| DL ≠ LD: twin jumps, commutator η(L − L−⊤)∇ℓ, boost test eφ | Transforming the Hamiltonian before an approximate solve (similarity renormalisation group, in-medium, transcorrelated, Lee–Suzuki); bosonic Bogoliubov transformations; preconditioning of variational wavefunctions; covariant fission-path searches | Most promising. New: a method for choosing the transformation through the gate, with the bound eφ, the norm of the inverse map that bounds any similarity transformation, as the acceptance test | Lean-verified: the twin-step and commutator identities, the orthogonal case and the boost relations. Implemented: boosts preserve distances to 7 × 10⁻¹⁵ |
| Two phases: small → frozen large through a control set | Emulators (eigenvector continuation, reduced basis); the light-to-heavy-nuclei gap | Untested; a direct test is defined | Lean-verified: the monad laws of the effect monad, all-or-nothing runs, persistence of fixed points and the memory bounds. Implemented on a 32-billion-parameter language model: the CEAS program and the fact update through the control set. Not yet run on nuclear data |
| Hecke operators on an arithmetic group; joint Hecke–Laplace eigenfunctions | A kernel basis on the model's own geometry for fitting nuclear data | Untested. Hecke operators exist only for arithmetic groups, and the GRAIL builds use Schottky groups. On a 32-billion-parameter model, the coset matrices of T₂ and T₃ on six heads cost about 0.2%; no automorphic kernel was formed, so they did not act as Hecke operators on the model | Lean-verified: the mode count (Weyl–Selberg, λ/12). Implemented: coset matrices ran on six heads of a 32-billion-parameter model |
Fusion
Magnetic-confinement fusion gives two close links: Greene’s residue, exact for the classification of periodic field lines, and the poloidal-angle twin, exact before the truncation of the series.
Inertial-confinement fusion overlaps with stockpile stewardship, so work stays on magnetic confinement and open fusion-energy data.
New physics: theoretical results
| Logarchéon piece | Fusion counterpart | Status | Verification |
|---|---|---|---|
| DL ≠ LD: twin maps | Changing the poloidal angle of the Fourier series of a flux surface leaves the surface unchanged, and the loss unchanged up to the truncation of the series, yet gradient optimisers (DESC, SIMSOPT) converge differently; angle choice by spectral condensation is known (Hirshman–Meier; Hirshman–Breslau) | A twin before truncation; a constant shift commutes with descent. Choosing a non-constant angle through the gate is new and testable | Lean-verified: the general commutator identity. Implementation in a stellarator optimiser not yet run |
| Two phases: small → large through a control set | Disruption predictors trained on small tokamaks transfer poorly to ITER or SPARC | Strong and testable | As for two phases on nuclei; not yet run on tokamak data |
| Langlands and Hecke operators | A kernel basis on the model's geometry for plasma data, as for nuclear data | Untested; the same limits as for nuclear data | As for the Hecke kernel basis on nuclei; not yet run on plasma data |
What the dimensional analysis gives
- ◆One regularity. Whether the ratio Tf/Tc of the temperature at which a Gaussian model space reaches its lowest level to the pairing temperature is the same across nuclei computed in one fixed model space. It is a regularity of statistical spectroscopy, not a new constant.
- ◆Twin preconditioning. Symmetry-generated transformations of variational and ab initio solves, accepted through the gate with the bound eφ; the measured quantity is the number of steps K to fixed accuracy.
- ◆Fusion methods. The gated choice of a poloidal angle in stellarator optimisation, and Greene’s residue in the group-attention module with the Lefschetz sum in the trace store.
Placed in nuclear relations, the parameters of the designs acquire dimensions, as h acquired the dimension of angular momentum through E = hν. The dimension names the known quantity to compare a parameter with; it does not make the parameter a new constant. The table gives the new results; Propositions 14 and 15 come from DL ≠ LD. Definitions and proofs →
New physics: theoretical results
| Parameter | Acquired dimension | Physical reading | Value | Status | Verification |
|---|---|---|---|---|---|
| Δτ = ln|a/a′|/E₁ (Proposition 14) | MeV⁻¹ | Imaginary time a twin saves; the twin leaves the gap E₁ unchanged at an invariant minimum | 1/E₁ = 0.0505 MeV⁻¹ per unit of ln|a/a′| for ⁴He | Theorem | Proved and Lean-verified. Implemented: on 20 random invariant quadratics the spectra agree to 7 × 10⁻¹³ and the saved steps match ln|a/a′|/E₁ asymptotically; on an attention model with the exact twin Q → e^s Q, K → e^{−s} K the late convergence rates agree to 2% |
| ½(max Ei − min Ei) over a twin orbit (Proposition 15) | MeV | Lower bound on the truncation error of a nuclear calculation, from the truncated solves alone | Computed per model space | Theorem; with the variational principle the complete spread | Proved and Lean-verified, in both forms. Implemented: Woods–Saxon well in a truncated oscillator basis, three levels; the bound holds and its variational form is met with equality to two digits |
| φ*/ζF | 1 | Optimal gated twin rapidity over the Fermi rapidity ζF = arsinh(pF/m) | ζF = 0.276 (symmetric matter), 0.346 (neutron matter) at saturation | Candidate regularity: new if the ratio is the same across nuclei | ζ_F interval-checked (0.276; 0.346). The ratio φ*/ζ_F is not yet computed |
| Tf/Tc | 1 | Ratio of the freezing temperature Tf of a fixed model space to the pairing critical temperature Tc = 0.567 Δ | Computed nucleus by nucleus in one model space | Candidate regularity, untested: new if the ratio is the same across nuclei | Lean-verified: the entropy-deficit and heat-capacity identities, the freezing condition, T_f = 0.209 σ for ²⁸Si and the BCS coefficient; interval-checked numerically. The ratio across nuclei is a theoretical computation in fixed shell-model spaces, not yet run |
Symbols used in these tables
| Symbol | Meaning | Dimension |
|---|---|---|
| ħ, kB | Reduced Planck constant and Boltzmann constant, both set to 1, as is the speed of light | — |
| Ej, Ē, σ | Energy of state j; centroid (mean) energy of a model space; width (standard deviation) of its energies | MeV |
| N | In formulas, the number of states of a model space; in "nucleus (Z, N)", the neutron number | 1 |
| Z | In formulas, the partition function Σj e−βEj; in "nucleus (Z, N)", the proton number | 1 |
| β; 1/β | Inverse temperature; temperature | MeV⁻¹; MeV |
| S, Smax | Entropy of the ensemble; its maximum, ln N | 1 |
| C | Heat capacity, β² Var(E) | 1 |
| T | Gap score, β/βc = βσ/√(2 ln N); T = 1 at the critical inverse temperature | 1 |
| Tf; βc | Temperature σ/√(2 ln N) at which a Gaussian model space reaches its expected lowest level; critical inverse temperature √(2 ln N)/σ | MeV; MeV⁻¹ |
| Tc, Δ | Critical temperature of the pairing transition; pairing gap | MeV |
| A; a; E*; ρ | Mass number; level-density parameter (≈ A/8 MeV⁻¹); excitation energy; level density | 1; MeV⁻¹; MeV; MeV⁻¹ |
| ħω | Spacing of the major nuclear shells, ≈ 41 A−1/3 MeV | MeV |
| sj; q, k | Attention score of key j; query and key vectors of an attention head | as mapped by ι |
| c | Curvature of the hyperbolic (lifted) attention of a head | MeV⁻² when q is a momentum |
| κ; ζ | Dimensionless curvature, c⟨|q|²⟩, equal to the mean squared rapidity ⟨ζ²⟩ under the dictionary; rapidity, v = tanh ζ | 1 |
| Λ | Mass scale c−1/2, set to m by the dictionary | MeV |
| p, m, v, γ | Three-momentum; mass; speed as a fraction of the speed of light; Lorentz factor 1/√(1 − v²) | MeV; MeV; 1; 1 |
| φ; eφ | Rapidity of a boost of the parameters; its error bound, of the form of a Doppler factor | 1 |
| θ; ℓ; ∇ℓ | Model parameters; loss (a variational energy under ι); its gradient | 1; MeV; MeV |
| L; L−⊤ | Twin map, an invertible linear map of the parameters; its inverse transpose | 1 |
| η; K; τ | Step size of gradient descent; number of steps; imaginary time Kη | MeV⁻¹; 1; MeV⁻¹ |
| H; G; E₁ | Hamiltonian; quantum geometric tensor; lowest excitation of the same quantum numbers | MeV; 1; MeV |
| ℍⁿ; Γ; Γ\ℍⁿ | Hyperbolic space of dimension n; a discrete group of its isometries; the quotient space | — |
| PSL(2,ℤ) | The modular group, the standard arithmetic group acting on ℍ² | — |
| ΔΓ; λ; r | Laplacian on Γ\ℍⁿ; its eigenvalue; spectral parameter, λ = 1/4 + r² on ℍ² | 1 |
| h, g | Test function of the trace formula and its Fourier transform | — |
| γ (in sums over Γ); ℓγ; ℓ₁ | A group element or closed geodesic; its length; the shortest length | 1 |
| tγ; ε₀ | Period of the orbit of γ; energy per unit geodesic length | MeV⁻¹; MeV |
| Tp; λp | Hecke operator at the prime p; its eigenvalue | 1 |
| μ; b, α; Δt; γRG | Renormalisation scale; constants of the depth law; renormalisation-group time per layer; scaling exponent of the temperature coupling | MeV; 1; 1; 1 |
| ι | Dictionary from model variables to physical quantities | — |
| EG; E₀; kT | Gamow energy of a fusion reaction; energy of the Gamow peak; thermal energy | MeV; keV; keV |
| S₀; ΔE | Configurational entropy of a protein's misfolded states; width of their energies | 1; energy |
| Tg; Tfold; Tphys | Glass temperature; folding temperature; physical temperature 1/β | energy |
Where the designs enter nuclear science
Nuclear science divides into fourteen branches: eight of basic science and six of application. The table is an index of where the designs and their quantities enter each branch. Details →
| Branch | Where the designs enter |
|---|---|
| B1 Nuclear structure and rare isotopes | Gap score and Tf of shell-model spaces (statistical spectroscopy); spacing-ratio statistics through the spectral face; periodic-orbit shell structure through the trace store |
| B2 Nuclear astrophysics | 1/β in statistical reaction rates; κ of dense matter |
| B3 Hot and dense QCD (heavy ions) | 1/β of hadron ensembles; κ and rapidity distributions; Lorentz-exact distances |
| B4 Hadron structure (Electron-Ion Collider) | Lorentz-exact distances on four-momenta |
| B5 Fundamental symmetries and neutrinos | Spacing statistics: a broken symmetry mixes sequences and changes the spacing ratio |
| B6 Nuclear theory and computation | Twin preconditioning of variational and ab initio solves (DL ≠ LD, the eφ bound); γRG only if the depth law holds |
| B7 Nuclear data | Spacing-ratio flags for missing levels or mixed assignments; Ψ abstention in model fits; provenance by hash chain |
| B8 Accelerators and detectors | Lorentz-exact distances in event reconstruction; κ of beam kinematics |
| B9 Fission and nuclear energy | Covariant fission-path search with the inertia tensor as metric; spacing-ratio flags and Ψ abstention in evaluations |
| B10 Fusion energy | Greene's residue in the group-attention module, with the Lefschetz sum in the trace store; the gated choice of a poloidal angle; κ of runaway electrons |
| B11 Isotopes for medicine and industry | 1/β in statistical production |
| B12 Nonproliferation and safeguards (open science) | Ψ: identification with abstention for attribution; spectral signatures through the spectral face |
| B13 Gravitation and multimessenger astrophysics | κ of neutron-star matter; image sums and the Selberg zeta for BTZ quasinormal modes, a known closed form used as validation |
| B14 Radiation effects and space nuclear systems | κ of relativistic particles; 1/β in statistical transport |
Energy: fission power, inertial fusion energy and the stars
In fission, the capture and fission cross sections of fuel nuclei in the unresolved-resonance region come from average resonance parameters (Hauser–Feshbach; Gilbert–Cameron), and delayed-neutron yields and decay heat from β-decay strength; the designs enter through spacing-ratio flags on evaluated levels, Ψ's abstention in model fits, and fission-path searches that use the inertia tensor as metric. Barrier penetration needs the WKB or instanton action, which the twin step does not compute. In plasmas and stars the rate weight exp(−βE − √(EG/E)) is the standard product of the thermal factor and the Gamow factor: EG = 1.182 MeV for D–T, with the Gamow peak at 30.9 keV for kT = 10 keV; EG = 493 keV for p + p, with the peak at 6.0 keV for kT = 1.34 keV. These are background numbers, not results of the designs. In neutron stars the mean squared rapidity of nucleons is κ = ⟨ζ²⟩ = 0.046 in symmetric matter at saturation density and 0.073 in neutron matter, and 0.094 and 0.147 at three times saturation.
Programmes for each branch (as of 20 October 2026)
The direct entry for a small business is the SBIR/STTR route; other programmes take a university or national-laboratory partner. Details →
| Branch | Programme | Entry |
|---|---|---|
| All branches | U.S. Department of Energy SBIR/STTR, including the Genesis Mission topics | Small business (Phase I guideline $200k, maximum $250k) |
| All branches | Genesis Mission rounds (teams of two of: national laboratory, industry, university; 20% cost share for a for-profit); next round planned for FY27 | Lead or team member |
| B1–B5 | DOE Office of Science, Nuclear Physics (AI/ML topic of the open call) | With a university or laboratory partner |
| B6 | DOE Advanced Scientific Computing Research (scientific machine learning, AI for science) | With a partner; DOE SBIR/STTR |
| B7, B12 | Nuclear Data Interagency Working Group call (Nuclear Physics and Fusion Energy Sciences; uncertainty quantification) | All applicant types |
| B8 | DOE Nuclear Physics SBIR/STTR topics (electronics, accelerators, instrumentation) | Small business |
| B9 | Genesis Mission challenge on nuclear energy | Team member |
| B10 | DOE Fusion Energy Sciences (AI/ML for fusion and plasma science) | With a partner; DOE SBIR/STTR |
| B11 | DOE Isotope R&D and Production | With a partner; DOE SBIR/STTR |
| B13 | DOE High Energy Physics (computational and theoretical research) | With a partner; DOE SBIR/STTR |
| B3, B8, B10, B14; mathematics and AI | Air Force Office of Scientific Research open BAA FA9550-26-S-0001: High-Energy Radiation-Matter Systems (nuclear and plasma physics); Computational Mathematics; Dynamical Systems and Control; physics-enhanced machine learning; Quantum Information Sciences; open until superseded | Industry and small businesses |
| B6, B8, B14; mathematics | Army Research Office and Army Research Laboratory BAA W911NF-23-S-0001 for foundational research; open to 20 November 2027 | For-profit organizations, large and small |
Other fields that share the mathematics
The mathematics of Logarchéon is not specific to nuclear physics. Five fields share several of its tools at once; eight more share one or two. The closest match is protein folding: there the critical point of CEAS corresponds to the glass temperature of the energy landscape, and the gap score is Tg/Tphys, glass over physical temperature. Proposition and details →
| Field (several tools) | Shared mathematics |
|---|---|
| Protein folding and biophysics | Random energy model of folding (Bryngelson–Wolynes, 1987): the critical point of CEAS corresponds to the glass temperature Tg, and the gap score is Tg/Tphys; energy landscapes, inverse design, causal inference |
| Condensed matter on hyperbolic lattices | Fuchsian groups and automorphic functions in hyperbolic band theory (Kollár et al., 2019; Maciejko–Rayan, 2021); Selberg trace formula; Hecke operators for arithmetic lattices |
| Quantum chaos, black holes, quantum information | Random-matrix statistics; SYK and JT-gravity matrix models; Hecke operators in AdS₃; quantum codes on hyperbolic surfaces (Breuckmann–Terhal, 2016) |
| Particle physics and QCD | Chiral random-matrix theory of the Dirac spectrum (Shuryak–Verbaarschot, 1993); Lorentz-equivariant analysis; contour deformations for the sign problem; renormalisation group |
| Quantum chemistry | Similarity transformations before an approximate solve (DL ≠ LD); neural-network wavefunctions with natural-gradient preconditioning; level statistics |
| Field (one or two tools) | Shared mathematics |
|---|---|
| Machine-learning theory | Spin-glass loss landscapes; random-matrix spectra of trained weights; natural gradient; hyperbolic representations of hierarchies (Nickel–Kiela, 2017) |
| Quantitative finance | Cleaning of correlation matrices at the Marchenko–Pastur edge (Laloux et al., 1999); causal inference |
| Wireless communications | Random-matrix capacity of multi-antenna systems (Telatar, 1999) |
| Network science | Hyperbolic geometry of complex networks and greedy routing (Krioukov et al., 2010); Ramanujan graphs as optimal expanders |
| Neuroscience | Ising models of neural populations (Schneidman et al., 2006); criticality and avalanches; hyperbolic geometry of hippocampal representations (Zhang et al., 2023) |
| Optimisation | Temperature control in simulated annealing; Boltzmann sampling; spin-glass formulations of combinatorial problems |
| Cosmology | Compact hyperbolic universes and their eigenmodes; random energy landscapes |
| Number theory and cryptography | Hecke operators, zeta functions, elliptic curves; post-quantum constructions |
Encrypted-in-use deployment
The seed architecture includes a λ-secure runtime for deployments requiring encrypted-in-use computation. Buyer-held keys. Policy-gated interfaces. No canonical plaintext during execution. Operates on the buyer’s hardware or within the buyer’s cloud tenancy.
V1 — λ-native models
Geometry built into model architecture and training dynamics from the ground up. Maximal integration. Principled semantics. Tighter control of canonicalisation. For long-lived sovereign AI assets where architectural integrity is non-negotiable.
V2 — Exported wrapper (NN/LLM)
Adoption-first path. Wraps existing models and runtimes without complete re-architecture. Fastest path to pilots. Reduces reusable plaintext exposure in in-use pipelines via constrained interfaces and protected representations.
V3 — VM/OS/runtime posture
Extends the same non-canonical in-use discipline to OS/VM/runtime boundaries for general-purpose compute — not solely AI. Covers cloud instances, on-premises deployments, hypervisor surfaces, and complete artifact lifecycle control.
Why homomorphic encryption (FHE), secure multi-party computation (MPC) and trusted execution environments (TEEs) fall short
FHE: 10³–10⁶× overhead, impractical for large neural pipelines. MPC: communication latency dominates at scale. TEEs: shift trust to vendor firmware, not zero-trust. All reintroduce plaintext through telemetry, caches, or debug steps.