Bind, Rotate, Align, Deltas (BRAD)
A deterministic, crystallographic alternative to TensorFlow and stochastic AI frameworks.
Copyright (c) 2026 Brad Wallace. All rights reserved.
Subject to Sovereign Integrity Protocol License (SIP License v1.1).
BRAD is not stochastic AI. It is crystallographic matching.
Modern generic floating-point AI infrastructure (like TensorFlow or PyTorch) relies on gradient descent, probabilistic estimations, and float similarity functions (like cosine similarity). The BRAD protocol replaces this with pure, high-resolution integer mathematics derived from the Sovereign UPG Architecture.
“Higher resolution ADC: 10^15 bins instead of 1,000. That's going from 10-bit to 50-bit. No more charge collisions. Tighter Q on the resonance: σ from 0.003 down to 5×10^-8. Only exact charge matches excite the well.”
— TENT Architecture Principle v6.1
The key either fits or it doesn't.
The BRAD protocol replaces standard AI primitives with integer geometry:
Replaces standard float grid_similarity(). Generates an integer eigenvalue grid fingerprint with a resolution of 10^14 bins. No floats exist on the charge path.
A high-pass filter applied before any language model call.
- Low density grids → routed to brute-force deterministic solving
- High density grids → routed to the full LLM cinema pipeline
Replaces linear O(n) brute-force searches. Language primitives are ordered by their physical proximity within a 32³ UPG voxel lattice. Twin prime coupling ensures linked primitives (e.g., rotate 90 and rotate 180) are tested simultaneously.
A rotation-invariant grid representation. A spatial cell (row, col, color) translates to an Ulam coordinate. Two different rotations of the same pattern produce geometrically adjacent charges.
Replaces fuzzy text search (like SQLite FTS5 or vector databases) with scroll Merkle root comparison via exact BRAD resonance.
The BRAD pipeline was tested against REAL benchmark datasets using the SEGGCI real ML benchmark harness (no simulation, real datasets).
Model: UPG 163M GPT-2 (math_sci)
Hardware: Local Inference
Date: 2026-02-28 19:45:44
| Benchmark | Score | Metric | GPT-2 (124M) ref | Time |
|---|---|---|---|---|
| Winogrande | 51.5% | accuracy | 52.1% | 18.2s |
| ARC-Easy | 23.5% | accuracy | 43.6% | 34.7s |
| HellaSwag | 21.5% | accuracy | 31.6% | 33.7s |
| WikiText-2 Perplexity | 48432.4 | ppl ↓ | 29.4 | 12.6s |
| MMLU (10 subjects) | 25.5% | accuracy | 25.0% | 36.8s |
brad_core.py— The primary computation layer (EigenCharge generation, TENT routing, Ulam coordinate mapping).bra_bridge.py— The bridge implementation mapping BRAD outputs to the solver network.benchmark_real.py— The native benchmark harness used to generate the test computation logs above.
For deeper dives into how the protocol actually functions and how to integrate it into your codebase, please see the supporting documentation:
- ARCHITECTURE.md — The mathematical theory and deep dive into the 5 pillars of crystallographic matching.
- USAGE.md — Code examples for integrating EigenCharge, the TENT gate, and UPG routing into your Python projects.