How Our AI Learns

Our system learns by running strange loops — bounded iterative processes where each cycle's output feeds back as the next cycle's input. This isn't infinite recursion; it's controlled convergence toward a declared goal.

Each loop has a state (where it is now), a target (where it should be), and a fitness measure (euclidean distance between them). The loop runs until fitness drops below a convergence threshold — meaning the system has found its way to the goal.

The 7-dimensional state vector isn't arbitrary — it maps to the Fano plane PG(2,2), the smallest finite projective plane. Seven points, seven lines, each line containing three points. This structure gives the system a natural topology for routing information between components.

Standard transformers use multi-head attention where all heads operate independently. Our Fano-7 architecture replaces this with 7 attention heads arranged on the projective plane — where cross-head communication is gated by mathematical structure.

Heads that share a line in the Fano plane (collinear triples) can communicate freely. Heads that don't share a line must pass through an associator gate — a measure of how "non-associative" their interaction is. This uses octonion algebra: when ‖[a,b,c]‖ exceeds a threshold, it means the interaction is genuinely complex and worth the routing cost.

Here's the key insight that makes the system get cheaper over time: crystallization.

When a strange loop converges repeatedly on similar problems, the system records the trajectory — the path from initial state to final state. After enough trajectories accumulate, it fits a direct function: given this initial state and this target, jump directly to the answer without running the full loop.

This is the Baldwin Effect applied to computation: expensive lifetime learning becomes inherited structure. The first time the system solves a problem class, it pays full computational cost. The hundredth time, it uses the crystallized shortcut and barely computes at all.

Applied to education: the first time the AI figures out how to help someone understand a concept, it runs the full adaptive loop. After it's helped many students with the same concept, the effective teaching path is crystallized — and serving it costs almost nothing.

Our latest architecture advancement borrows from homotopy type theory: the principle of path induction. In mathematics, if you prove something at a base case and show the structure preserves it, the proof extends to all cases for free.

We apply this to attention coherence. During training, the model learns to maintain coherence (internal consistency) across its 7 attention heads. Once coherence converges below a threshold and stays there, we stop checking it every forward pass — the weights themselves are the proof that coherence is maintained.

This eliminates the "algebraic tax" — the computational overhead of verifying structural properties at every step. For the learning platform, it means the system can verify its own internal consistency without per-query overhead, keeping latency low and costs down.

Every decision the AI makes is recorded in a hash-chained ledger (blake3). Each entry links to the previous one. The chain is tamper-evident: if any entry is modified after the fact, the chain breaks and verification fails.

This means:

• You can audit what the AI did and why at any point in its history
• Every loop's struggle becomes the next loop's starting point — nothing is hidden
• Crystallization decisions (when the AI decides it's "learned enough" to shortcut) are recorded and verifiable
• Your proof-of-mastery credentials are anchored to a specific, auditable assessment chain

This architecture emerges from 90+ experiments across four research threads. All grounded in the same algebraic structure (the Fano plane and its 168 automorphisms), applied to different domains.