Persson et al. 2026

Statistical foundations of LLM-based A/B testing: a surrogacy framework. Persson, E., Schultzberg, M., and Ankargren, S. (2026). arXiv:2606.17165.

Relationship to this project — Formal causal frame

Formalizes the assumptions and calibration under which effects on LLM outcomes identify effects on human outcomes; prior validation does not verify validity for a new intervention. This project supplies a design-side registered example: coarse marginal validation left the incentive-response estimand weakly constrained and the microstructure unidentified.