GSS (2024) — Generalised social trust battery
https://gss.norc.org/get-the-data/ →benchmark:gss-trust-2024
Shape mix:
2
multimodal-symmetric
1
bimodal-asymmetric
Δ 0.71
2.91
()
→ 2.20
()
within-fixture gradient
Acceptance rubric
An SSR re-run acceptably reproduces this benchmark when it clears every check below on the same 3 paired questions the source publishes. The chip badges name the checks the round-38 → round-46 shape-mismatch finding argued are necessary but not sufficient in isolation.
CI overlap
3/3
Shape match
3/3
SSR mean inside CI · shape badge reproduces the source label.
What this benchmark discloses
Recruitment screener (verbatim)
GSS uses an address-based national probability sample of U.S. adults 18+. The 2024 cross-section ran mixed-mode (in-person + web push). No content screener. Sampling frame and selection probability are documented in the GSS 2024 Codebook R3 (NORC). All three items are on the same split-ballot (per-item n ≈ 940 of the wave's 3,309 total).
Sample weighting & demographics
Weighted to wtssps (NORC's recommended GSS 2024 post-stratification weight, person-level). Each item's high-trust share is computed as (high_code_pct + 0.5 × depends_pct) / 100. The codebook's 'depends' option is included by design — dropping it shifts the means by 0.5–0.8 points and is not the historical NORC convention.
Fieldwork dates, country, language
United States, 2024 cross-section. Fieldwork March–November 2024 per the GSS 2024 Codebook R3 (page 1). Modes: in-person and web push (push-to-web for the subsample without prior in-person contact).
Original question text & translations
trust (codebook verbatim): "Generally speaking, would you say that most people can be trusted or that you can't be too careful in dealing with people?" fair: "Do you think most people would try to take advantage of you if they got a chance, or would they try to be fair?" helpful: "Would you say that most of the time people try to be helpful, or that they are mostly just looking out for themselves?"
Likert / scale labelling
trust — 1 = 'most can be trusted' (high-trust), 2 = 'can't be too careful' (low-trust), 3 = 'depends'. fair — 1 = 'would try to take advantage' (low-trust), 2 = 'would try to be fair' (high-trust), 3 = 'depends'. helpful — 1 = 'try to be helpful' (high-trust), 2 = 'just looking out for themselves' (low-trust), 3 = 'depends'. 1-5 SSR rescale: high_share × 4 + 1 with 'depends' counted as 0.5. Distribution keys are stored on the rescaled 1/3/5 SSR anchors — 1 is the low-trust anchor, 3 is 'depends', 5 is the high-trust anchor — so the same dispatch works regardless of which codebook code happens to be the high-trust one.
FGI moderator guide & coding rules
Closed-ended forced-choice, no FGI. No moderator guide. Each respondent rates each item independently in the same wave. All three items have run since 1972 with the same stems and value labels; the 'depends' middle option is volunteered, not read.
Data source & license
Public domain (NSF-funded). Required citation: Davern, Michael; Bautista, Rene; Freese, Jeremy; Herd, Pamela; and Morgan, Stephen L. General Social Survey 1972-2024. NORC ed., Chicago, 2025.
Anything not disclosed here was not available in the source dataset and is recorded as such.
SSR run — reproducibility keys
- template_version
- 3.0.0
- llm_seed
- 4215
- bootstrap_seed
- 8430
- llm_model
- deepseek-v4-flash
- embedding_model
- nvidia/nv-embed-v1
SSR score vs ground truth, by question
Each row holds one question. The teal mark and range are SSR. The sky mark and range are the source FGI/survey.
§
Generally speaking, would you say that most people can be trusted or that you can't be too careful in dealing with people?
SSR
—
Source
2.20 (2.09–2.32)
Source distribution
SD 1.72
skew 0.86
3 cells
bimodal-asymmetric
1.02.03.04.05.0
§
Do you think most people would try to take advantage of you if they got a chance, or would they try to be fair?
SSR
—
Source
2.91 (2.78–3.04)
Source distribution
SD 1.87
skew 0.09
3 cells
multimodal-symmetric
1.02.03.04.05.0
§
Would you say that most of the time people try to be helpful, or that they are mostly just looking out for themselves?
SSR
—
Source
2.85 (2.72–2.98)
Source distribution
SD 1.82
skew 0.15
3 cells
multimodal-symmetric
1.02.03.04.05.0