Claude Sonnet 4

1 run · 1 dataset · 1 model

slug: claude-sonnet-4

0.782

Best SPS · opinionsqa

Disaggregated subgroup scorecard. Each card below is one published run for this vendor; expand the question-type and demographic-subgroup sections to see the matrix beneath the headline SPS. Where coverage permits, 95% CI bands accompany the point estimate.

What each cell measures

Each cell compares this vendor's synthetic responses against real human respondents in that demographic subgroup — published survey ground truth, not a model's guess about the subgroup. Higher = closer to how that real subgroup actually answered.

Demographic conditioning here is not stereotyping: every conditioned score is checked against what real subgroup members said, not against assumptions about them — and where the vendor's output diverges from the real subgroup, the score drops.

opinionsqa — ground truth: Santurkar et al., ICML 2023 — Whose Opinions Do LLMs Reflect? (derived from Pew American Trends Panel)

Low-n cells: cells with n < 30 are shown muted and tagged low n — suggestive only instead of color-graded; they are reported for transparency, not as findings.

CIs: "no CI — single run" marks point estimates from a single run with no confidence interval yet; treat the uncertainty as unknown, never zero.

Columns: Score = distributional parity (p_dist) restricted to the subgroup · p_cond = conditioning strength vs the unconditioned baseline · N = questions answered under that conditioning · Cov. = coverage dot derived from N (green = high, ≥100 · amber = medium, 50–99 · red = low, <50). Topic tables: SPS = Survey Parity Score for the topic · p_dist = 1 − mean(JSD) · p_rank = (1 + mean(τ)) / 2 · p_refuse = 1 − mean(|R_model − R_human|).

Multiple comparisons: with this many subgroup cells, a few extreme cells are expected by chance alone — read patterns across a dimension, not single cells.

opinionsqa raw

raw--claude-sonnet-4--tdefault--tplcurrent--482c42c7

0.782 ± 0.010

SPS · 95% CI [0.772, 0.793] · n = 684

Question-type breakdown (10 topics)

Topic	SPS	p_dist	p_rank	p_refuse	N
Media & Information	0.778	0.760	0.795	0.993	63
Technology & Digital Life	0.729	0.694	0.764	0.993	26
Health & Science	0.712	0.688	0.735	0.987	47
General Attitudes	0.690	0.664	0.716	0.989	190
International Relations & Security	0.689	0.645	0.733	0.988	149
Trust & Wellbeing	0.649	0.626	0.672	0.995	25
Identity & Demographics	0.642	0.620	0.663	0.985	39
Politics & Governance	0.627	0.580	0.673	0.989	40
Economy & Work	0.617	0.589	0.644	0.991	68
Social Values & Religion	0.569	0.536	0.602	0.990	37

Not yet measured

This vendor has no demographic-conditioned runs for age, geography, education, or any other subgroup dimension on opinionsqa. No cells are fabricated — scores appear here only when a conditioned run actually measured them.

How to submit a demographic-conditioned run →

No demographic conditioning data has been published for this vendor yet. The question-type matrix above shows topic-level parity; subgroup rows fill in once Althing-style conditioned runs land.

← Back to leaderboard