BURT-IMMA

Matrix-Memory Equilibrium Propagation - a 13-layer architecture that forks the backprop lineage.

SnapKitty West / SNAPKITTYWEST - Evidence or Silence - 2026


Model Description

13-layer architecture. Learning rule is Matrix-Memory Equilibrium Propagation (MMEP): two-phase, locally computable, Hebbian-style update. No gradient tape. No weight transport.

Core memory cell (CIFG):

C_t = f_t * C_{t-1} + (1-f_t) * (v_t outer k_t)
Component Description
CIFG Matrix Memory Full d x d state, sum-inversion retrieval
MMEP Learning Rule Two-phase local update, no backprop
Entropy-Constrained Router H(alpha) <= 0.20 nats (sovereign_entropy.tex)
Spectral Projection Contractive relaxation guarantee
SmoothLeaky Activation 4-axiom formal specification
Boolean Perceptron Actors Huntington postulate-verified routing
7 CUDA Kernels sm_86 (RTX 3080)
Lean 4 Formalization Machine-checked proof ledger

Training

  • Hardware: RTX 3080 sm_86, 7 custom CUDA kernels
  • Learning rule: MMEP (no Adam, no SGD)
  • Corpus: SNAPKITTYWEST/sovereign-training-corpus
  • Weights: Architecture + code released; pretrained weights pending

Papers - 11 Total (8 Zenodo + 3 In-Repo LaTeX)

File Lines Venue Title
gdr_kernels.tex 460 SC / MLSys GDR-9: Formally Verified Generalized Delta Rule Kernel Stack Across Nine ISAs
liquidops_kernel.tex 566 PLDI / ICFP LiquidOps: Refinement-Verified NAND-Canonical Kernel for Data-Plane ISA Compilation
sovereign_entropy.tex 394 FM / CAV Sovereign Entropy: Formally Verified 0.20-Nat Bound for Deterministic Agent Outputs

GDR-9 - Nine-ISA kernel stack (CUDA, NASM, RISC-V, ARM, WASM, APL, Haskell, Lean 4, P4). Generalized delta rule + drain invariants. Cross-ISA equivalence proved in Lean 4. Target: SC/MLSys.

LiquidOps - NAND-canonical verified compiler. LiquidHaskell refinement types enforce correctness end-to-end. P4 backend. Target: PLDI/ICFP.

Sovereign Entropy - Lean 4 machine-checked proof: routing entropy H(alpha) <= 0.20 nats. SPARK Ada contract + ERE gate. Governs BURT-IMMA router and sovereign-memory-twin verdict threshold. Target: FM/CAV.

Zenodo (8 papers, published 2026-07-01) - PIRTM, octonions, Coxeter/Weyl, PH-DAE, EmojiScript, Goldilocks, Resonance disclosure, DMZ F2 decomposition. DOIs in sovereign-engine-v2 README.


License

Apache-2.0

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train SNAPKITTYWEST/burt-imma