Simulatte Research — 2026

The numbers.

Three studies. Two countries. One question: can a synthetic population match how real people actually think?

US Population Accuracy
10%
2.3pp from the human self-consistency ceiling. Pew Research Center benchmark.
India Population Accuracy
46%
First company to replicate India's political landscape at population scale.
vs. Average LLM
1.0×
Closer to the human ceiling than the average large language model. 10 LLMs tested.
Simulatte International Benchmark — United States — 2026
57.6%

2.3 percentage points from the theoretical human ceiling.

We took 15 published Pew Research Center questions and ran them through a synthetic US population of 60 personas. The result: 88.7% distribution accuracy, cohort-adjusted — sitting just 2.3pp from the ceiling that even humans can't surpass.

60 personas 15 questions Pew Research Center ground truth Reproducible
Read technical report → Audit data ↗
Simulatte Personas vs Pew — Accuracy Racing to Human Ceiling
100% 75% 50% 25% HUMAN CEILING 91.0%
Gap to ceiling
2.3pp
Total improvement
+31.1pp
Simulatte India Benchmark — 2026
46.2%

The first company to replicate India's political landscape at population scale.

Using Sarvam's proprietary AI infrastructure for language and cultural grounding, we rebuilt India's political landscape from scratch — 40 personas, 22 architectural sprints. Ground truth: Pew Research Center + CSDS-Lokniti survey data, the gold standard for Indian public opinion.

40 personas 22 sprints Pew + CSDS-Lokniti ground truth Sarvam infrastructure
Read technical report →
From baseline to benchmark
Simulatte India Benchmark (current)85.3%
Unoptimized baseline46.2%
Total gain
+39.1pp
Sprints
22
Personas
40
Simulatte vs the LLMs — 2026
1.0×

Closer to the human ceiling than any LLM tested.

We ran 10 large language models against the same India Pew benchmark. 5,878 SHA-256 verified API calls. The gap between Simulatte and the best available LLM is not close.

10 LLMs tested 5,878 verified API calls SHA-256 checksums
Read technical report → Audit data ↗
Performance vs Human Ceiling — India Benchmark
Human self-consistency ceiling (Iyengar et al.)
91.0%
Simulatte
85.3%
GPT-4o — best LLM tested
75.6%
Gemini
≈44%
Average LLM — 10 models
≈17%
India Pew + CSDS-Lokniti ground truth · 5,878 SHA-256 verified API calls
Methodology

How accuracy is measured

Distribution accuracy = 1 − Σ|real_i − sim_i| / 2. Identical formula to the UC Berkeley synthetic population study, enabling direct comparison. Human ceiling of 91.0% sourced from Iyengar et al. (Stanford). Every result reproducible via public GitHub repository.

Read accuracy report → US study — full technical report → GitHub: raw data and audit artifacts ↗