Overview
Study Reconstruction Pipeline
The model is not fed “a persona.” It is fed standardized tuples: respondent, condition, question, answer.
Record Unit
Tuple Anatomy
Each row is a tiny experimental scene. Fine-tuning teaches the model to map that scene onto an answer distribution, not just generate plausible prose.
Hover cells to see what gets preserved when messy study materials are flattened into training rows.
Corpus Shape
210 Studies as a Heatmap
Hot cells mark studies that are harder to standardize or simulate: more arms, more outcome scales, or denser response volume.
Scale
Counts and Domain Mix
Orders of magnitude matter here: participant counts, condition counts, and outcome counts do not live on the same scale.
Evaluation
How the Tuned Models Pull Away
Distribution View
Shape Beats Fluency
The win condition is not “sounds human.” It is matching how human answers spread across bins and how that spread shifts under treatment.
Human
Prompted
Fine-tuned
Interpretation
Plausibility vs Causal Faithfulness
The episode’s core warning lives in this quadrant: better distribution matching can still leave the simulator below the bar for causal trust.
Pressure Tests
Where Errors Hide
One chart asks whether treatment direction survives. The other asks whether aggregate improvement still leaves hot subgroup disparities.