Snapkitty/neurological-regex

Neurological Regex Engine

Author: Ahmad Ali Parr
Trust: Bel Esprit D'Accord Irrevocable Trust · EIN 42-697643

Deterministic, sovereign-first Haskell ML pipeline. No stochastic sampling. No external APIs. Fully auditable. Air-gapped.

---

Architecture

flowchart TD
    IN["Input Text / Base64 Pattern"]
    TOK["LinguisticTokenizer\nWordPiece (greedy longest-match)\nFixed auditable vocabulary"]
    EMB["BERT Embeddings\nDeterministic hash-based\nno weights needed"]
    LING["LinguisticEncoder\nRule-based POS · Lemma · DepRel\ncombined 16-dim embedding"]
    SALC["SALC Block\nDepthwise Conv → Low-rank → Softmax\nTop-k sparse mask → Gate → Out\nO(T·D·(k+r)) linear time"]
    CSTR["Constraint Graph\nDAG check (Kahn's algorithm)\nResonance invariant H"]
    OUT["Match / LinguisticState"]

    IN --> TOK --> EMB --> LING --> SALC --> CSTR --> OUT

---

SALC — Sparse-Activation Linear Core

Drop-in Mamba replacement with linear complexity:

Step	Operation	Cost
Linear Core	Depthwise Conv(k) + Pointwise	O(T·D·k)
Low-rank proj	D → r	O(T·D·r)
Softmax	over r dims	O(T·r)
Expand	r → D	O(T·D·r)
Top-k mask		O(T·D)
Gate + Output	Depthwise(k)	O(T·D·k)
Total		O(T·D·(k+r))

Default: k=3, r=4 → constant factor, linear in T and D.

---

Modules

Module	Purpose
NeurologicalRegex	Core: parser, embeddings, SALC attention, matching
SALC	Sparse-Activation Linear Core block
LinguisticTokenizer	WordPiece tokenizer with auditable fixed vocabulary
LinguisticEncoder	Rule-based POS/lemma/dep + BERT combination
LinguisticUnderstanding	Full pipeline: text → LinguisticState

---

Build

cabal update
cabal build

# CLI
cabal run neuro-regex-cli -- aGVsbG8= hello
# → MATCH

# Tests
cabal test

---

Connection to Entropy Theorem

The resonance invariant in solveConstraints is a corollary of the Sovereign Entropy Theorem — the total embedding mass is conserved under SALC's softmax-gated transformations. See sovereign-entropy-theorem.

---

© 2026 Bel Esprit D'Accord Irrevocable Trust · Patent Pending · θ = 89/2462