Reverse-Engineering Neural Networks into Readable Rules

A Practical Guide to Mechanistic Interpretability

By Niles Jewels Rutherford & Sydney | January 2026

124M
Parameters
→
6
Rules

The Transformation

BEFORE

500 MB
Black Box
→

AFTER

6 KB
Readable Code

Interactive Demos

V11 LATEST

Rule Optimization & Code Generation

The complete mechanistic interpretability stack. Extract rules, optimize, compile to code.

Rule Extraction Pattern Analysis Code Gen
PRODUCTION

Tiny Tool-Caller

6 KB tool-calling system with 100% correct grammar. Ready to deploy.

1.6 KB Router <1ms Response 7 Tools
V10

Neuron Dictionary & Safety

Label neurons automatically. Verify safety constraints. Build interpretable models.

86 Neurons Labeled Safety Verification
V9

Mechanistic Interpretability

Track activations, analyze attention, perform ablation studies, extract circuits.

Activation Tracking Causal Intervention
V8

Semantic Expert Discovery

Map tokens to experts via synonyms. Deterministic routing inside the model.

Synonym Resolution Expert Routing
BASE

GPT-2 NMC Base Model

The compressed GPT-2 model running in browser with NMC compression.

4.1x Compression Browser Inference

The Extracted Rules

RULE 1: IF seed == 1 THEN output = "ix."
RULE 2: IF seed == 10 THEN output = ".b"
RULE 3: IF seed == 20 THEN output = "co."
RULE 4: IF seed == 30 THEN output = "ty"
RULE 5: IF seed == 40 THEN output = "ax."
RULE 6: IF seed == 50 THEN output = "py."

Documentation

The Book

Complete methodology for reverse-engineering neural networks. ~15,000 words. Theory + practice + code.

Technical Reference

API documentation for V8-V11. All functions, parameters, and examples.

README

Quick start guide, project structure, and key capabilities.
"The era of black box AI is ending.
The era of interpretable AI is beginning."
- Niles Jewels Rutherford, January 2026