Research-driven tools for AI evaluation and alignment diagnostics

Reproducible evaluation suite implementing diagnostic tests from three LLM behavior research papers:

• Phi Eval — Epistemic pathology detection (overconfidence metrics)
• DI Eval — Delegated introspection measurement (reflective thought migration)
• OT Bench — Observer-time diagnostics (temporal consciousness tests)

All experiments run in < 2 minutes with mock mode for reproducibility.

Our evaluation frameworks operationalize theoretical work in:

• Epistemic virtue theory (Zagzebski, Roberts & Wood)
• Speech-act theory (Austin, Searle)
• Phenomenology of time (Husserl, Merleau-Ponty, Sartre)
• RLHF alignment research (Christiano et al.)

All manuscripts are currently under peer review. Repositories contain implementation code and evaluation frameworks only.

To develop faithful, minimal-compute experiments that:

1. Operationalize theoretical claims about LLM behavior
2. Enable reproducible diagnostics without proprietary data
3. Distinguish registration from constitution in machine cognition
4. Map epistemic pathologies to measurable metrics

We welcome:

• Additional model implementations (local, open-source)
• Extended question sets and evaluation scenarios
• Real user studies to complement simulated dialogues
• Multi-language support

All contributions must maintain theoretical fidelity to source papers.

All projects: MIT License

Bentley DeVilling — Course Correct Labs Boulder Creek, CA coursecorrectlabs.com Bentley@CourseCorrectLabs.com

© 2025 Course Correct Labs