Nº 17b · Replication human-baseline replication map · 80 years of social science

Sixty years of behavioral science. One arc.

43
published studies passing design filters
0.73
mean rank correlation with human outcomes
0 / 52
competitors making this claim · 2026 dossier
The classical literature found the stated/revealed gap. The LLM-as-subject literature began closing it. Subconscious productionizes the closure. Each milestone below links to the paper.
  1. 1957
    Classical · stated/revealed identified
    Festinger & Carlsmith · Stanford
    Cognitive consequences of forced compliance.
    Paper ↗ Bibliography ↗
  2. 1974
    Classical · choice modeling · load-bearing for our method
    Daniel McFadden · Berkeley
    Conditional logit analysis of qualitative choice behavior — the random-utility framework Subconscious uses.
    Paper ↗ Nobel lecture ↗
  3. 1979
    Classical · gap formalized
    Kahneman & Tversky · Econometrica
    Prospect theory: an analysis of decision under risk.
    Paper ↗ Bibliography ↗
  4. 2022
    LLM-as-subject · feasibility
    Argyle, Busby, Rasmussen, Gubler, Bail · BYU + Duke
    Out of One, Many — using language models to simulate human samples.
    Paper ↗ Bibliography ↗
  5. 2024
    LLM-as-subject · scale
    Park, Zou, Shaw, Hill, Cai, Morris, Willer, Liang, Bernstein · Stanford
    Generative agent simulations of 1,000 people.
    Paper ↗ Bibliography ↗
  6. 2025
    Production · the closure shipped
    Subconscious · causl.io
    Reviewed replication set — 43 published studies passing design filters · mean rank correlation 0.73.
    Full corpus ↗ Discord ↗
MethodSynthetic respondents drawn from a U.S.-population trait distribution. Each study replayed under the original protocol. Effects compared at the effect-size level, not the surface answer.
FidelityMean rank correlation is 0.73 across the 43 published studies passing design filters. The measured human ceiling is 0.959.
Why this mattersThe dossier surveys 52 commercial competitors. Zero make a human-baseline replication claim. The substrate — not the chatbot — is the moat.
More research go.subconscious.ai/data ↗
Join the discussion Discord ↗