The annals
v0.2 makes the assembly less theatrical and more truthful: minds with covariance, crowds with census weight, numbers with humility, and an agora where opinions meet.
v0.2
Why: real minds are not five independent dials; inter-trait correlations preserve the joint personality structure that marginal matching can lose.
van der Linden et al. 2010, meta-analytic Big Five inter-trait correlations, N=144k; Williams et al. 2026, arXiv:2601.15755
v0.2
Why: each summoned panel now begins from the population ledger - age pyramids, education tables, and income-class shares - before it speaks.
UN WPP 2024; ACS 2023; Census-of-India education tables; national income-class shares
Why: IPF raking brings the synthetic sample back to its target margins, while Kish effective sample size tells how much signal remains after weighting.
Pew-style survey weighting; iterative proportional fitting; Kish effective sample size
v0.2
Why: every verdict now carries a 95% interval, so certainty must be earned rather than implied by a single gilded number.
Bootstrap 95% confidence intervals on v0.2 verdict outputs
Why: each result receives a B or C tier from the evidence available, with variance-collapse diagnostics against human purchase-intent dispersion.
Maier et al. 2025, arXiv:2510.08338; SSR reaches roughly 90% of human test-retest reliability
v0.2
Why: when ordinary prompting narrows the chorus, probability-weighted sets of possible reactions recover 1.6-2.1x of the diversity lost to mode collapse.
Zhang et al. 2025, arXiv:2510.01171
Why: Friedkin-Johnsen dynamics over a small-world network show how word of mouth can move the verdict, including early warnings for cascade risk.
Asch-line conformity literature; Han et al. 2026, arXiv:2601.05606; confidence-weighted distributed pooling
The verdict is still a screening signal. It is now clearer about the minds it summons, the crowd it represents, and the uncertainty it must confess.
Return to the method