The central question
When an AI system appears to learn or remember, which observable changes support that description? The lab studies tasks, memory records, information retrieval, and the validity of measurements used to evaluate those behaviors.
The method
The approach separates model capability, tools, memory, and measurement instruments. Comparable test conditions help identify whether a failure came from the model, execution, available data, or the evaluation itself.
Status and limits
This is an independent research line with working manuscripts. Internal drafts are not peer-reviewed publications. A completed run or an interesting result does not by itself prove a hypothesis.