A synthetic sample is a language model asked to answer a survey as if it were your respondents. This one hit the real survey's average exactly, the first number any buyer checks. That badge stays green no matter what you do next. Run each test on the same sample and watch what the matched average was hiding.
Faults hidden under this matched mean: 0 of 5
Illustrative sample, not one recorded prompt. Every fault shown is the documented behaviour in real synthetic-respondent studies: variance collapse (Bisbee et al., 2024), RLHF narrowing (Kirk et al., 2024), sycophancy (Sharma et al., 2024), WEIRD skew (Atari et al., 2023; Santurkar et al., 2023). Frozen time is a mechanism argued by analogy, not a measured survey failure. Fix status recorded 2026-08-18; a Webflow embed cannot call a model.