SpecHO v2

Unified AI Text Detection & Model Fingerprinting

Detect AI-generated text with 98.6%+ accuracy. Identify which model wrote it.

---

What It Does

• Binary Detection — Human or AI?
• Model Attribution — GPT-4, Claude, Gemini, Llama, etc.
• Model Fingerprinting — Build profiles over time
• AURORA Integration — Trust verification

Quick Start

from src import SpecHO

specho = SpecHO("data/fingerprints")

# Analyze text
result = specho.analyze(
    "Your text here",
    model_id="gpt-4"
)

# Verify model identity
verification = specho.verify(
    "Unknown text",
    claimed_model="claude-sonnet-4"
)

Fast detection (< 1ms):

from src.fingerprint import LightweightClassifier

clf = LightweightClassifier()
is_ai, confidence = clf.predict("Some text")

Architecture

flowchart TD
    A[Input Text] --> B{Tier Selection}
    B -->|Fast| C[Layer E - 6D]
    B -->|Medium| D[Layers E+A+B - 26D]
    B -->|Full| E[All 45D]
    C --> F[Classification]
    D --> F
    E --> F
    F --> G[Human/AI + Model ID]

Loading

45D Feature Layers

Layer	Dims	Features
A	5	Semantic trajectory
B	15	SpecHO echo patterns
C	12	Epistemic + transitions
D	7	Syntactic rhythm
E	6	Lightweight structural

Tiered Extraction

Choose your speed/accuracy tradeoff:

Tier 1 — Fast

• Layers: E only (6D)
• Latency: < 1ms
• Accuracy: 98.6%
• Use: High-volume screening

Tier 2 — Standard

• Layers: E + A + B (26D)
• Latency: ~50ms
• Accuracy: ~99%
• Use: Normal analysis

Tier 3 — Full

• Layers: All 45D
• Latency: ~100ms
• Accuracy: > 99%
• Use: Deep verification

Performance

Validated on 464-sample corpus:

Metric	Old (24D)	New (45D)
Accuracy	94.4%	> 98%
Human Recognition	72.4%	> 92%
False Positive Rate	27.6%	< 8%

Key fix: Restored Layer B (15D echo features) missing from old system.

Project Structure

Click to expand

Core (src/)

Path	Description
fingerprint/	Layer extractors
├─ lightweight.py	Layer E (6D)
├─ layer_b_echo.py	Layer B (15D)
├─ trajectory.py	Layer A (5D)
├─ epistemic.py	Layer C.1 (6D)
├─ transitions.py	Layer C.2 (6D)
├─ syntactic.py	Layer D (7D)
└─ unified_45d.py	Combined
echo_engine/	Core analysis
preprocessor/	spaCy pipeline
clause_identifier/	Detection
scoring/	Aggregation

Web (web/)

Path	Description
index.php	Frontend
api/	Python backend
static/	JS/CSS

Data (data/)

Path	Description
models/	Trained classifiers
fingerprints/	Reference DB
validation/	Test results

Tools (tools/)

Path	Description
train_specho.py	Training
ab_test.py	A/B testing
visualize_*.py	Visualization

Key Files

File	Purpose
src/fingerprint/lightweight.py	Fast 6D detection
src/fingerprint/unified_45d.py	Full extraction
src/specHO_integrated.py	High-level API
web/index.php	Web frontend
specho_cli.py	CLI tool

Installation

Minimal (Tier 1)

pip install numpy

Full (All Tiers)

pip install numpy scipy scikit-learn spacy sentence-transformers
python -m spacy download en_core_web_sm

CLI Usage

# Analyze text
./specho_cli.py "Your text here"

# From file
./specho_cli.py -f document.txt

# JSON output
./specho_cli.py --json "Your text"

# Verbose
./specho_cli.py -v "Your text"

API Reference

SpecHO Class

class SpecHO:
    def analyze(
        text: str,
        model_id: str = None
    ) -> AnalysisResult

    def verify(
        text: str,
        claimed_model: str = None
    ) -> VerificationResult

    def identify(
        text: str
    ) -> VerificationResult

LightweightClassifier

class LightweightClassifier:
    def predict(text: str) -> Tuple[str, float]
    # Returns: ("HUMAN" | "AI", confidence)

    def predict_proba(text: str) -> Dict[str, float]
    # Returns: {"human": 0.8, "ai": 0.2}

License

MIT

---

_{Built with cognitive fingerprinting technology}