Adaptive World Model Evaluation

When is the model unreliable? How should it adapt? How long should we trust it?

3 Research Papers
100% Open Source
R Resonance Signal

The Resonance Framework

All components share a common signal: resonance (R) — the agreement between learned model predictions and physics-based priors.

R = exp(-||f_θ(x) - Φ(x)||² / σ²)

High R → Trust the model | Low R → Adapt, blend, or rebuild

🎯

Detection

When is the model unreliable? RBD detects distribution shift via model-physics resonance.

⚖️

Adaptation

How should it adapt? HHA+MIT provides stress-based control with Adapt/Freeze/Rebuild decisions.

⏱️

Horizon

How long to trust predictions? ARH+DH adapts evaluation horizons and discriminates noise from drift.

Publications

Overview

Companion Note

Unified View of the Resonance Framework

A synthesis document explaining how the three components interact and their theoretical foundations.

RBD

Resonance-Based Detection

Distribution Shift Detection via Model-Physics Agreement

R-weighted blending between model and physics predictions. When model and physics disagree, trust physics. When they agree, trust the model. Wrong physics gracefully degrades to baseline.

  • +50.2% OOD improvement over baseline
  • Graceful degradation with wrong physics
  • No retraining required
HHA+MIT

Homeostatic Control & Model Invalidity

Stress-Based Adaptation with Adapt/Freeze/Rebuild Taxonomy

An agent that regulates learning via stress signals. HHA provides continuous control, MIT elevates stress into discrete epistemic decisions with 100% noise specificity.

  • Three-regime taxonomy: Adapt/Freeze/Rebuild
  • 100% noise specificity (0 false reconstructions)
  • Panic Freeze protective mechanism
ARH+DH

Adaptive Horizons & Failure Diagnosis

Temporal Dynamics of World Model Evaluation

ARH adapts evaluation horizons based on confidence. DualHorizon discriminates noise from drift using sign structure analysis, validated on 25 real-world TCPD datasets.

  • Confidence-modulated horizon (4-20 steps)
  • 76% discrimination on TCPD benchmark
  • Noise vs drift diagnosis

Open Source

# Central entry point
git clone https://github.com/TheCause/resonance-framework.git

# Individual components
git clone https://github.com/TheCause/resonance-based-detection.git      # RBD
git clone https://github.com/TheCause/homeostatic-hamiltonian-agent.git  # HHA+MIT
git clone https://github.com/TheCause/temporal-evaluation-framework.git  # ARH+DH

# Run benchmarks (example: HHA+MIT)
cd homeostatic-hamiltonian-agent
pip install -r requirements.txt
python run_all_benchmarks.py

All code is MIT licensed. Contributions welcome.

Contact

For questions about the research or collaboration opportunities: