Tag
This paper examines the gap between reliability and construct validity when using LLMs as coding instruments for theoretical constructs, and proposes grain calibration as a method to decompose constructs into clause-level components for more valid measurement.