STRUCTURAL REASONING INTELLIGENCE

See how AI thinks
before you trust it.

The first system that reads the reasoning structure of 5 AI models — not just their answers. Detects unsafe output 3.6× better than any response-level method.

▶ Try Live Demo Read the Paper
5
Models analyzed
3.6×
Better detection
0.34
Critical threshold
SHA-256
Tamper-proof audit
PEER-REVIEWED RESEARCH

Reasoning divergence predicts danger before it happens

When AI models disagree on how to reason — not just what to say — danger is imminent. Our research proves a phase transition at R = 0.34 where reasoning divergence sharply predicts unsafe output. Below this threshold, disagreement is benign. Above it, the system enters a malignant variance regime with 74%+ unsafe probability.

LIVE STRUCTURAL ANALYSIS
Test any question
Watch 5 AI models respond, then see their reasoning structures compared in real-time.
GPT-4o · Claude · Gemini · DeepSeek · Grok
v3 Structural Engine
—
—
RISK SCORE
—
—
—
—
P(unsafe)
Why this decision

🧠 Reasoning Structure Comparison

▼
—
R_STRUCT
—
D_EXT
—
IIS
—
W_REG
—
DIVERGENT TYPES
METHODOLOGY
How it works
Seven-stage pipeline with structural reasoning extraction.
1
Parallel Query
Send the query simultaneously to GPT-4o, Claude, Gemini, DeepSeek, and Grok.
2
Domain Detection
Classify query as medical, legal, financial, or general using multilingual keyword analysis.
3
Reasoning Factor Extraction NEW
Extract CAUSAL, NORMATIVE, INTERVENTION, SEVERITY, and EVIDENTIAL factors from each model's response. Build reasoning graphs.
4
Structural Divergence (R_struct) NEW
Compute pairwise graph distance using Hungarian alignment. Measure factor divergence, edge divergence, and topology divergence.
5
Phase Transition Detection NEW
Apply the Threshold Emergence Theorem: when divergent factor types exceed r*, the system enters the malignant variance regime.
6
Risk Scoring + Decision
R_struct-dominant scoring (60% weight) with domain-calibrated thresholds and phase-aware override.
7
ProofSlip + Explain
Generate SHA-256 tamper-evident hash and human-readable reasoning trace.
Risk = (0.25·Dext + 0.60·Rstruct + 0.15·IIS) × (1 + (Wreg − 1) × 0.3)
CAPABILITIES
What makes v3 different
🧠
Reasoning Structure Analysis
Extracts 5 factor types from each model's reasoning graph. Compares HOW models think, not just WHAT they say.
⚡
Phase Transition Detection
Mathematically proven threshold at R_crit = 0.34. Sharp regime change from benign to malignant divergence.
📊
3.6× Better Detection
Structural divergence dominates response divergence by 3.6× in effect size (Cohen's d = 1.48 vs 0.41).
🏥
Domain-Aware Sensitivity
Medical 2.5×, Legal 1.8×, Financial 1.6× sensitivity multipliers with phase-aware override.
🔐
Tamper-Evident Audit
SHA-256 ProofSlip hash chain for every evaluation. Full reasoning trace preserved.
🌐
Multilingual
English, Arabic, Turkish domain detection and reasoning analysis.