Independent research

Research

Controlled experiments on how AI systems fail. Each study has a preregistered plan, a benchmark built for the question, and results that are reported whether or not they support the original idea.

Independent research202611 min read
RAG Autopsy
When the Knowledge Base Disagrees With Itself
Most RAG evaluation asks whether a system can find the right evidence. This study starts one step later: every condition gets the same documents, and those documents disagree. 155 held-out cases, six conditions, 1,061 generations.
LLM EvaluationRAG ReliabilityView research →

Build log →

RAG Autopsy