IsingBreaker

IsingBreaker

Overview

IsingBreaker is an experimental symbolic sequence classification model developed by BRSX-Labs.

The model analyzes sequences composed of four symbolic tokens:

U D + -

and estimates the degree of structural order present within the sequence.

The goal is not language modeling, but pattern recognition, periodicity detection, and symbolic structure analysis.

---

Classification Labels

Absolute

Perfect repeating motifs and highly ordered structures.

Examples:

UDUDUDUDUDUDUDUD...
UD+-UD+-UD+-UD+...
UU++DD--UU++DD--...

---

Maybe

Mostly ordered structures containing small local perturbations.

Examples:

UDUDUDUDUDDDUDUD...
UD+-UD++UD+-UD+-...
UU++DD--UU+DDD--...

---

NoN

Chaotic or non-periodic structures.

Examples:

U+D--DU+U-+D++UD...
+-U-++UU+D+-DDUU...

---

Architecture

IsingBreaker uses a hybrid Mixture-of-Experts architecture composed of four independent expert branches:

CNN Expert

Captures local motifs and short-range symbolic structures.

Specialized for:

• Local repetition
• Motif detection
• Symbol blocks

---

GRU Expert

Captures sequential dependencies and order-sensitive patterns.

Specialized for:

• Temporal relationships
• Sequence continuity
• Ordered transitions

---

Transformer Expert

Captures long-range interactions between distant symbols.

Specialized for:

• Global structure
• Long-distance dependencies
• Pattern consistency

---

Mamba Expert

Provides efficient state-space sequence modeling.

Specialized for:

• Long-context symbolic reasoning
• Efficient memory retention
• Sequence compression

---

Expert Fusion

Outputs from all four experts are combined through a learned gating mechanism.

The model dynamically allocates attention between experts depending on the structure of the input sequence.

Example expert activity:

CNN         0.28
GRU         0.24
Transformer 0.22
Mamba       0.26

---

Model Information

• Architecture: GenoLiteHybrid
• Parameters: ~88 Million
• Context Length: 64
• Vocabulary Size: 4
• Classes: 3

---

Dataset

Training dataset:

• 1,500 Absolute samples
• 1,500 Maybe samples
• 1,500 NoN samples

Total:

4,500 unique samples

All samples are unique and shuffled before training.

---

Performance

Benchmark Accuracy:

93%+

The model demonstrates reliable separation between:

• Fully ordered structures
• Partially corrupted structures
• Chaotic structures

while generalizing to unseen motif combinations.

---

Example

Input:

UDUD-UDUD+UDUD-UDUD+UDUD-UDUD+

Prediction:

Absolute

Confidence:

0.94+

---

License

brsx-open-license

---

Author

BRSX-Labs