We systematically evaluate large language models by comparing their generated citations and attributions against critical textual editions. This process identifies systematic representation errors and hallucinated references or outcomes in complex philosophical and spiritual corpora.
Our methodology adapts established cybersecurity vulnerability assessment techniques to textual scholarship, focusing on data integrity and the verifiable provenance of knowledge representation within AI systems. This ensures objective, reproducible results.
Foundational Research Principles
Empirical Testing
Neutral Analysis
Open Data
Zero Hype
All evaluations are data-driven, focusing on quantifiable metrics of citation accuracy and attribution fidelity across diverse textual traditions.
We maintain strict objectivity, avoiding devotional bias or speculative interpretations. Our findings are presented without advocacy.
We aim to be open with datasets and evaluation protocols to foster transparency and independent verification.
Our analysis prioritizes critical assessment over promotional claims, addressing real-world implications for AI development and deployment.