Metadata-Version: 2.4
Name: entropy-predict
Version: 0.2.0
Summary: Predictive intelligence through agent-based population simulation
Project-URL: Homepage, https://github.com/exaforge/entropy
Project-URL: Repository, https://github.com/exaforge/entropy
Project-URL: Issues, https://github.com/exaforge/entropy/issues
Author: Devesh Paragiri
License-Expression: MIT
License-File: LICENSE
Keywords: agent-based,llm,population,predictive-intelligence,simulation
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.11
Requires-Dist: anthropic<1,>=0.77.0
Requires-Dist: networkx<4,>=3.4.0
Requires-Dist: numpy<3,>=2.0.0
Requires-Dist: openai<2,>=1.50.0
Requires-Dist: pydantic<3,>=2.9.0
Requires-Dist: python-dotenv<2,>=1.0.0
Requires-Dist: pyyaml<7,>=6.0.0
Requires-Dist: rich<14,>=13.0.0
Requires-Dist: scipy<2,>=1.14.0
Requires-Dist: typer<1,>=0.12.0
Provides-Extra: dev
Requires-Dist: pytest-asyncio>=0.24.0; extra == 'dev'
Requires-Dist: pytest>=8.0.0; extra == 'dev'
Requires-Dist: ruff>=0.14.0; extra == 'dev'
Description-Content-Type: text/markdown

# Entropy

[![CI](https://github.com/exaforge/entropy/actions/workflows/test.yml/badge.svg)](https://github.com/exaforge/entropy/actions/workflows/test.yml)
[![PyPI version](https://img.shields.io/pypi/v/entropy-predict.svg)](https://pypi.org/project/entropy-predict/)
[![Python](https://img.shields.io/pypi/pyversions/entropy-predict.svg)](https://pypi.org/project/entropy-predict/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

Predictive intelligence through agent-based population simulation. Create synthetic populations grounded in real-world data, simulate how they respond to events, and watch opinions emerge through social networks.

Not a survey. Not a poll. A simulation of collective human behavior.

## What It Does

You describe a population and a scenario. Entropy builds statistically grounded synthetic agents, connects them in a social network, and has each one reason individually about the event using an LLM. Opinions form, spread through the network, and evolve — producing distributional predictions you can segment and analyze.

```
entropy spec → entropy extend → entropy sample → entropy network → entropy persona → entropy scenario → entropy simulate
                                                                                                  │              │
                                                                                           entropy estimate    entropy results
```

## Install

```bash
pip install entropy-predict
```

Or from source:

```bash
git clone https://github.com/exaforge/entropy.git
cd entropy
pip install -e ".[dev]"
```

## Setup

```bash
# API keys (in .env or exported)
export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...

# Or for Azure OpenAI:
# export AZURE_OPENAI_API_KEY=...
# export AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com
# export AZURE_OPENAI_DEPLOYMENT=your-deployment-name

# Configure providers (openai, claude, or azure_openai)
entropy config set pipeline.provider claude      # Claude for population/scenario building
entropy config set simulation.provider openai    # OpenAI for agent reasoning
entropy config show
```

## Quick Start

```bash
# Build a population
entropy spec "500 Austin TX commuters who drive into downtown for work" -o austin/base.yaml
entropy extend austin/base.yaml -s "Response to a $15/day downtown congestion tax" -o austin/population.yaml
entropy sample austin/population.yaml -o austin/agents.json --seed 42
entropy network austin/agents.json -o austin/network.json -p austin/population.yaml --seed 42
entropy persona austin/population.yaml --agents austin/agents.json

# Compile and run a scenario
entropy scenario -p austin/population.yaml -a austin/agents.json -n austin/network.json -o austin/scenario.yaml
entropy simulate austin/scenario.yaml -o austin/results/ --seed 42

# View results
entropy results austin/results/
entropy results austin/results/ --segment income
```

### What Comes Out

Outcomes are defined per-scenario — categorical, float, boolean, or open-ended. You choose what to measure.

```
═══════════════════════════════════════════════════════════
SIMULATION RESULTS: austin_congestion_tax
═══════════════════════════════════════════════════════════

Population: 500 agents | Duration: 47 timesteps | Model: gpt-5
Stopped: exposure_rate > 0.95 and no_state_changes_for > 5

EXPOSURE
────────────────────────────────────────
Final exposure rate: 96.8%
Reasoning calls: 1,847
Average conviction: 0.64 (moderate-to-firm)

OUTCOMES
────────────────────────────────────────
commute_response (categorical):
  drive_and_pay          38%  ███████████████░░░░░
  switch_to_transit      24%  █████████░░░░░░░░░░░
  shift_schedule         19%  ███████░░░░░░░░░░░░░
  telework_more          12%  ████░░░░░░░░░░░░░░░░
  undecided               7%  ██░░░░░░░░░░░░░░░░░░

sentiment (float, -1 to 1):
  mean: -0.18  std: 0.41  min: -0.9  max: 0.7

willingness_to_pay (boolean):
  yes: 42%  no: 58%

protest_likelihood (float, 0 to 1):
  mean: 0.31  std: 0.28

SEGMENT: income
────────────────────────────────────────
< $50k:   drive_and_pay 22% | switch_to_transit 14% | protest 41%
$50-100k: drive_and_pay 40% | switch_to_transit 28% | shift_schedule 21%
> $100k:  drive_and_pay 51% | switch_to_transit 31% | telework_more 14%
```

Each agent reasoned individually. A low-income commuter with no transit access reacts differently than a tech worker near a rail stop — not because we scripted it, but because their attributes, persona, and social context led them there.

The scenario YAML controls what gets tracked:

```yaml
outcomes:
  suggested_outcomes:
  - name: commute_response
    type: categorical
    options: [drive_and_pay, switch_to_transit, shift_schedule, telework_more, undecided]
  - name: sentiment
    type: float
    range: [-1.0, 1.0]
  - name: willingness_to_pay
    type: boolean
  - name: protest_likelihood
    type: float
    range: [0.0, 1.0]
```

## How It Works

**Population creation** — An LLM discovers relevant attributes (demographics, psychographics, scenario-specific), then researches real-world distributions with citations. Agents are sampled from these distributions respecting all dependencies. A social network connects them based on attribute similarity with small-world properties — the network config (which attributes matter, what edge types exist, who influences whom) is generated by the LLM from the population spec.

**Persona rendering** — Each agent gets a first-person narrative built from their attributes. Relative traits are positioned against population statistics ("I'm much more price-sensitive than most people"). Generated once per population, applied computationally per agent.

**Two-pass reasoning** — Pass 1: the agent role-plays their reaction in natural language (no enum labels, no anchoring). Pass 2: a cheap model classifies the freeform response into outcome categories. This eliminates the central tendency bias that plagues single-pass structured extraction.

**Network propagation** — Agents share information through social connections. Edge types, spread modifiers, and decay control how opinions travel. Multi-touch re-reasoning lets agents update their position after hearing from multiple peers.

## Documentation

- **[CLI Reference](docs/commands.md)** — Every command with arguments, options, and examples
- **[Architecture](docs/architecture.md)** — How the system works under the hood
- **[Use Cases](docs/use-cases.md)** — What Entropy can simulate, with examples

## Development

```bash
pip install -e ".[dev]"
pytest                    # Run tests
ruff check .              # Lint
ruff format .             # Format
```

## License

[MIT](LICENSE)
