EVO Classification White Paper
EVO: A Neuro-Symbolic Classification Engine for Enterprise Data
Download White Paper (PDF)Abstract
We present EVO, a classification engine built on Radial Starfish (RSF) spiking neural dynamics that achieves state-of-the-art results on enterprise schema matching, COBOL copybook translation, and image classification — without GPU, LLM, pre-trained embeddings, or external dependencies.
The system uses a cybernetic feedback loop that improves with every human interaction, reaching 87% on the Goby enterprise benchmark (75 types, 1,187 sources) with only 104 human decisions, 99% on CopyBench COBOL classification (23 copybooks, 301 fields), and 90.2% on MNIST image classification. Privacy mode incurs zero accuracy cost.
Benchmark Results
Enterprise Schema Matching
87.3% accuracy on the Goby benchmark — 1,187 data sources classified with only 104 human answers. Less than 1 question per 10 sources.
COBOL Copybook Classification
99.0% accuracy — 298 out of 301 fields correct across 23 real and synthetic copybooks. Total classification time: 8 milliseconds.
Image Classification
90.2% on MNIST handwritten digits without GPU or backpropagation. 80.7% on Fashion-MNIST clothing items.
Privacy at Zero Cost
Structural mode matches full mode exactly. The system classifies data from character patterns alone — without reading actual values.
What Sets EVO Apart
- Universal classifier — One engine classifies enterprise data, COBOL fields, images, and radio signals from a single architecture
- Self-improving — The only classification system that gets smarter with every use. Process 100 sources, it barely needs you
- Zero GPU, Zero LLM — No API calls, no pre-trained models, no external dependencies. Runs on any laptop
- Deterministic — 100% identical classifications after save/load roundtrip. Same input → same output, every time
- Air-gapped — Runs entirely offline. No internet needed. The data never leaves the room
Download the Full White Paper
Explore the complete architecture, benchmark methodology, and results in detail.
Download White Paper (PDF)