Evidence and limits
Finding: An observation stream, retrieval, reflection, and planning produced more believable simulated behavior than ablated variants.
Impact: Watch reflection-linked retrieval and planning as an implementation pattern. Reject believability as evidence of factual correctness or workflow benefit.
Boundary: The target was simulated social believability, not truth, authority, or professional task performance. Next: which reflection outputs can be checked against source evidence before reuse?
Flags: simulation; believability outcome; model-judge and rater dependence
Record State: Catalog review recorded; not a performance endorsement
What was reviewed: Structure and relevance screened; not a full evaluation