🧬 μ-GPT — Eigenverse Language Model
572 Lean 4 theorems → neural architecture → trained weights
A GPT variant where every architectural choice is derived from machine-verified mathematics in the Eigenverse:
| Component | Source | Theorem |
|---|---|---|
| Activation: C(r) = 2r/(1+r²) | §5 coherence_le_one | AM-GM bound |
| MLP width: ⌊δS·dim⌋ | §7 silver_ratio | Self-similar expansion |
| 8 heads (Z/8Z) | §15 rotational_memory | 8-cycle orbit |
| sech(λ) q-gain | §10 Lyapunov duality | Bounded dissipation |
| μ-phase rotation | §15 Z/8Z binding | Head-orbit coupling |
| LR warmdown: C(r) | §5 coherence_symm | Palindrome schedule |
Architecture: Muon optimizer, RoPE, U-Net skip connections, logit softcap.
Not a metaphor. Every constant is Lean-verified. Every design choice has a proof.