Validated against published survey ground truth — every run public

Our synthetic crowd matches real-world polling at 93.3% parity.

This is the proof the engine works — 40 topics across 4 independent survey houses (Pew Research Center, Gallup, GSS, KFF), with items chosen to expose weaknesses rather than hide them: dead heats, near-consensus questions, high-uncertainty topics. The latest run is the first Grade A in this benchmark's history — and the run that scored a C is published right beside it, because every run ships to this ledger, misses included. The same validated engine is then re-grounded and calibrated for each market, so it reflects local opinion.

Loading benchmarks…

Parity
93.3%
vs published polling · 4 survey houses
Mean abs. error
3.36 pp
inside a professional poll's own margin of error
Pearson r
0.984
correlation to real polling
Grade
A
parity banded A–F
Topic-ranking fidelity
—
Spearman across 40 topics — same order as the real world
Distribution divergence
—
Jensen-Shannon (log2) · 0 = identical distributions
Latest run
latest validated run
400 agents/question

Calibrated for your market

The 93.3% above proves the engine on the hardest public benchmark. For each market we don't reuse it — we rebuild the crowd from local census (in Singapore, the CMIO ethnicity, religion and home-language distributions) and anchor it to local survey data, so it reflects local opinion rather than a US echo.

Per-market grounding sources — US: Pew Research; other markets: World Values Survey, Reuters Institute, IPS, and national statistics offices. Absolute-level calibration is tuned to each client's live data.

How we compare

Every enterprise competitor publishes a parity number. Most are self-reported, against private benchmark sets. Ours is measured against published polling and validated end to end.

VendorClaimed parityValidation
CrowdOS (this page)93.3%Measured on 40 survey items · versioned script + ledger · reproducible
Evidenza88–95%EY-validated, $50K+ per engagement
Artificial Societies95%self-reported
Synthetic Users85–92%Park 2024 method
Ditto92%EY-audited

Measured, not asserted

This is a committed benchmark — a versioned script, a versioned ledger, and a fixed question set — not a marketing figure. We run it against our production backend and walk qualified buyers through the full methodology and per-question results under NDA.

For the methodology overview, see the methodology page.