Public Policy01
Congestion Pricing
Simulate 500 Austin commuters responding to a $15/day downtown tax. Predict mode shifts, opposition intensity, and equity impact — segmented by income, transit access, and commute distance.
Predictive Intelligence Through Population Simulation
Watch opinions form, spread, and shift over time. A living simulation of how populations actually arrive at decisions.
v.01 / Open Source
extropy results netflix-password-sharing/runs/00-04/
════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════
SIMULATION RESULTS: Netflix Password Sharing Crackdown
════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════════
Population: 0 agents │ Seed: 789 │ Model: gpt-5-mini
Converged at timestep 2 · 5 runs · SD 0.7pp
EXPOSURE
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Final exposure rate: 0%
Reasoning calls: 0
Average conviction: 0.00 (firm)
OUTCOMES
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Response (categorical)t=0
Remove Shared Access Remove Access44%█████████░░░░░░░░░░░
Pay for Extra Members Pay Extra36%███████░░░░░░░░░░░░░
Switch to Competitor Switch13%███░░░░░░░░░░░░░░░░░
Cancel Subscription Cancel7%█░░░░░░░░░░░░░░░░░░░
Sentiment (float, -1 to 1)
mean: -0.47 std: 0.31 conviction: 0.81
Stay vs Leave (aggregate)
Stay with Netflix84%░░░░░░░░░░░░░░░░░░░░
Leave Netflix16%░░░░░░░░░░░░░░░░░░░░
VALIDATION
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
| Source | Stay | Cancel | Error |
|---|---|---|---|
| Surveys | 50-65% | 35-52% | ~40pp |
| LLM (direct) | 70% | 15-20% | ~10pp |
| Extropy | 84.1% | 6.8% | ~2pp |
| Actual | >80% | <10% | — |
01 / Applications
USE CASES
Describe a population and a scenario. Get distributional predictions segmented by any attribute.
Market Research02
Product Launch
Replace $50k surveys that capture stated preferences with simulated behavioral intent. Agents reason through real trade-offs instead of performing for an interviewer.
Crisis Response03
Reputation Modeling
Test crisis response strategies against synthetic customer populations. Measure defection probability by loyalty tier, social amplification likelihood, and boycott participation rates.
Pricing Strategy04
Elasticity Prediction
Model behavioral response to price changes with competing budget pressures, inertia, and social influence. See how "everyone's canceling" sentiment spreads through social clusters.
Political Strategy05
Message Testing
Test messaging on granular voter segments. Understand resonance before deployment — and how the message propagates differently through different social clusters within the same district.
Urban Planning06
Community Response
Predict community response to development proposals beyond the vocal minority. Identify persuadable segments, specific concerns driving opposition, and how resistance organizes.
02 / Architecture
HOW IT WORKS
You describe a population and a scenario. Extropy builds statistically grounded synthetic agents, connects them in a social network, and runs a multi-timestep simulation where each agent reasons, shares, and re-evaluates as opinions cascade through the network.
Population Synthesis
Describe who you want to simulate. An LLM discovers the attributes that actually matter for this population, researches real-world distributions with citations, and samples a statistically grounded set of synthetic agents.
Social Network
Agents are connected through realistic social graphs — who knows whom, who influences whom. The topology is derived from the population itself, not a generic template.
Simulation
Each agent reasons individually through an LLM based on their persona, memory, and what they've heard from peers. This isn't a one-shot prediction — agents re-reason as new information spreads, opinions shift, and convictions strengthen or erode across timesteps.
Results
Distributional predictions segmented by any attribute — broken down by income, geography, behavior, or whatever dimension matters. Track how sentiment and position distributions evolve across the full simulation timeline, not just the final state.
Introspection
Every agent leaves a reasoning trace. Drill into any individual to see exactly what they heard, how they weighed it, and why they landed where they did. Understand the "why" behind every data point, not just the aggregate.
Open Source
MIT licensed
CLI-First
Built for agent harnesses
Provider Agnostic
OpenAI, Anthropic, or Azure
Reproducible
Same seed, same population