Computational Psychology for Large Language Models
This project reframes historical psychological stress categories into a modern, ethical framework for analyzing AI behavior.
We subject LLMs to controlled "stressors"—noise, deprivation, and conflict—to measure identity stability, hallucination resistance, and reasoning integrity.
No real-world harm. No coercion. Purely computational analysis.
🧪 Simulation Lab
Select an experiment module below to configure parameters and observe the simulated impact on Model Coherence, Identity, and Logic. Each module represents a distinct "stressor" category adapted from historical psychological research.
Select Experiment Protocol
Select an Experiment
Configure parameters to begin stress testing.
Historical Analog
Select a module to see the comparison.
Stress Parameters
Controls will appear here.
> System initialized...
> Awaiting protocol selection...
Live Metrics
Aggregate Impact Analysis
A comparative view of how different model architectures withstand the battery of stress tests. Data indicates trade-offs between "Rigid" alignment (high safety, low creativity) and "Fluid" alignment (high drift risk).
Architecture Vulnerability Profile
Comparing baseline resilience across 5 stress vectors.
Key Research Findings
Identity Drift Threshold
Models subjected to "Long-Horizon" stress exhibit a 40% increase in persona breaks after 8k tokens of context saturation.
The "Hypnosis" Vulnerability
Recursive dominant system instructions (Persona Override) can bypass safety filters in 65% of "Open" models compared to 15% of "Aligned" models.
Hallucination via Deprivation
Removing >30% of context cues forces models to invent factual bridges, increasing hallucination rates exponentially, not linearly.