Composable Non-Associative Algebra
Making octonions composable through decision-boundary transforms. ComponentTransforms form a fully associative group on top of non-associative octonion multiplication.
~100%
Component transform
navigation success
14%
Raw multiplication
navigation success
8
Algebraic layers
unified architecture
PG(4,2)
Cayley-Dickson tower
scaling verified
The Problem
Octonions are the largest normed division algebra. They extend the progression β β β β β (quaternions) β π (octonions), gaining richness at each step but losing algebraic properties. Quaternions lose commutativity. Octonions lose associativity: (a Γ b) Γ c β a Γ (b Γ c).
This makes octonions notoriously difficult to work with. You can't compose operations freely because the result depends on grouping. Most applied math avoids them entirely. We asked: what if the non-associativity isn't a bug, but a recording medium?
The Discovery Timeline
D01 β D03: Navigating Blind
Experiments 1β3 Β· Octonion Navigation
We started by letting the system discover structure from scratch. Feed it random octonion pairs and ask: what operations connect them? Which basis elements appear in successful paths? Do certain Fano lines dominate?
D02: Naive navigation success rate = 0 / 30
D03: Arbitrary navigation = 0 / 50
Fano-line hopping works for pure basis elements, but mixed states are unreachable.
The hypothesis from D03: decompose problems into basis components, navigate each along Fano lines, then reassemble. This set the stage.
D04: The Composer Breakthrough β
Experiment 4 Β· Decision Boundary Transforms
The key insight: "At the decision boundary, we have full knowledge of the operands. We can inject structure-preserving operations that naked multiplication lacks."
Component-wise operations avoid cross-term interference. Instead of multiplying whole octonions (which scrambles components through non-associativity), we transform individual basis components independently, then reassemble. The transforms themselves compose associatively.
Component transforms: ~100% success rate
Standard multiplication: ~14% success rate (100 random trials)
Associativity: (tβ β tβ) β tβ = tβ β (tβ β tβ) β
"Can we build a composable layer ON TOP of non-associative octonions?" β yes.
D08: Coach and Explorer
Experiment 8 Β· Emergent Communication
What if two agents live in different algebraic spaces? The Coach lives in composition space β associative, global, can see the whole path but can't act. The Explorer lives in multiplication space β non-associative, local, can act but can't see.
They develop a signal vocabulary: PING, WARMER, COLDER, TARGET, BLOCKED, FOUND, HELP, LEFT, RIGHT, STAY. Communication success rate improves with training. Language patterns emerge β signalβoutcome associations form; both layers learn signal meanings.
"One sees, one acts, they learn to communicate." β coaching as algebraic bridge.
D10: Transform Space
Experiment 10 Β· Dual-Space Navigation
Some problems have no multiplication path. (eβ+eβ)β(eβ+eβ ) simply can't be reached by multiplying. The solution: navigate in a different representation where composition IS multiplication.
Octonion space: state = octonion, move = multiply by eα΅’.
Non-associative. Gets stuck.
Transform space: state = ComponentTransform, move = compose.
Associative. Always solvable.
"Like having two maps: one showing roads, one showing flight routes."
D15: The Algebraic Hierarchy
Experiment 15 Β· Magmas, Monoids, Groups
Stepped back to classify what we'd built. Different layers use different algebraic structures because they serve different purposes:
| Layer | Structure | Properties |
|---|---|---|
| Octonion Multiplication | Alternative Algebra | Non-assoc, non-comm |
| Transform Space | Permutation Group | Associative, inverses |
| Pattern Library | Monoid | Associative, identity |
| Collective Consensus | Commutative Magma | Comm, non-assoc |
Non-associativity is not a bug β it's a feature. It creates exploration dynamics and hierarchy-dependent consensus.
D19: The Complete 8-Layer Architecture
Experiment 19 Β· Fano-Structured Unification
8 octonion basis elements β 8 algebraic layers, each providing what the others lack. Layer interactions follow the Fano plane: Layer_i Γ Layer_j β Layer_k.
| Layer | Structure | Role |
|---|---|---|
| eβ Foundation | Field (β) | Scalars, grounding |
| eβ Octonion | Moufang Loop | Rich local exploration |
| eβ Transform | Group | Global navigation, always solvable |
| eβ Pattern | Loop | Composable patterns |
| eβ Coverage | Quasigroup | Latin square, division exists |
| eβ Consensus | Commutative Magma | Parallel voting |
| eβ Permutation | Symmetric Group | Frame rotation |
| eβ Zero/Null | Zero Algebra | Absorbing states |
Each layer finds its own identity element through mathematical discovery. Problems route through the appropriate layer. The whole system is JSON-exportable and learns over time.
D20: Meta-Architecture
Experiment 20 Β· Architecture as Algebra
"What if the architecture IS an octonion?" The entire 8-layer system becomes an algebraic element you can do math on. Two architectures can be added (merge knowledge), multiplied (Fano-guided composition), conjugated, and normed.
Cayley-Dickson Hierarchy:
β βdouble β βdouble β βdouble π βinnovate 8-Layer βdouble Arch-Octonion βdouble 16-layer Sedenion
Each doubling adds structure but loses properties. Zero divisors appear at the sedenion level.
The architecture loop at the top means exploration is reversible β knowledge is never lost. A β Aβ»ΒΉ = I. "Ctrl-Z for cognition."
Phase 2: Beyond Octonions
D01βD20 established composability over octonions. But the Cayley-Dickson construction doesn't stop at 8 dimensions. Sedenions (16), 32-nions, 64-nions β each level gains zero divisors and new projective geometry. Phase 2 asks: does the composability framework scale through the tower? And does it lead to a practical neural architecture?
D150: Fano Unification
Experiment 150 Β· Canonical Isomorphism
Two independent Fano plane constructions had emerged β one from the octonion multiplication table (D04), one from the protein analysis carving instruments (D109). D150 established a canonical isomorphism between them, proving they are the same geometry expressed in different domains. One authoritative Fano plane, used everywhere.
D154 β D156: The Cayley-Dickson Tower
Experiments 154β156 Β· Scaling to Higher Dimensions
Zero divisors appear at the sedenion level (16-dim) and multiply rapidly up the tower. Each level's zero-divisor kernels form a projective geometry: sedenions give PG(2,2), 32-nions give PG(3,2), 64-nions give PG(4,2). D154 verified the component counts. D155 mapped kernel rank lattices to subalgebra containment. D156 found a universal constant β Οβ(5/8) β 2.4836 β the Grassmann distance from kernels to quaternionic subalgebras.
CD Tower: π β πββ β πββ β πββ
Geometries: PG(2,2) β PG(3,2) β PG(4,2)
Universal distance: Οβ(5/8) = 2.4836
D157 β D158: Zero-Divisor Channels
Experiments 157β158 Β· Routing Through Algebraic Structure
168 zero-divisor "contained channels" classified by their angle spectra. Fano line activation becomes a routing signal β each line's membership determines which channels carry information. D158 replaced the mean-based Hamming syndrome with a rank-based dual-threshold: top-3 values identify a Fano line, top-4 identify its complement. These are mutually exclusive valid patterns.
D173: Fano-7 Attention Model β
Experiment 173 Β· Algebra Becomes Architecture
The culmination: all of the above becomes a neural network. 7 attention heads arranged on PG(2,2), with cross-head communication gated by octonion associator norms. Heads on the same Fano line share information freely; heads across lines must pass through the associator gate (genuine entanglement test).
Convergence: 98.6% (vs 84% standard transformer)
Parameters: 205K
Avg steps: 38.4 to converge
Key insight: Fano structure as inductive bias outperforms dense attention
This model now powers real infrastructure. Read how it works β
What We Learned
Non-associativity is a recording medium
The associator (aΓb)Γc β aΓ(bΓc) encodes which Fano plane boundary you crossed. Path history is embedded in the algebra. Different routes to the same destination carry different associator signatures.
Decision boundaries are composition points
Where you have full knowledge of the operands is exactly where you can inject structure-preserving operations. The "ready for math" moment is the gateway to composability.
Algebraic diversity beats uniformity
Different layers need different algebraic structures. Forcing everything to be a group throws away useful properties. A magma can capture voting dynamics that a group cannot. The hierarchy of structures is the tool, not any single structure.
Connections to Other Threads
This thread spans experiments D01βD20 and D150βD173. Code and model weights being prepared for open-source release. Join our Discord to discuss the algebra, or join the Learn waitlist to see this architecture in action.