A market-type-aware AI evaluation platform for Polymarket. The system classifies prediction markets into reasoning categories, builds domain-specific context pipelines (adapters) for each type, and runs multiple AI agents against historical market states using a replay engine.

shared/         Shared dataclasses — single source of truth for all types
pipeline/       Workstream A: data fetching, market classification, adapters
engine/         Workstream B: agents, replay engine, position manager, metrics
app/            Workstream C: orchestration entrypoint, Streamlit dashboard
data/cache/     Pre-fetched market JSON (written by prefetch script, read at runtime)

The two integration points between workstreams:

# pipeline/__init__.py
def build_context(slug: str, replay_timestamp: datetime) -> tuple[AdapterContext, MarketSnapshot]: ...

# engine/__init__.py
def run_evaluation(context: AdapterContext, market: MarketSnapshot) -> list[AgentEvalResult]: ...

# 1. Install dependencies
pip install anthropic httpx streamlit pandas python-dotenv

# 2. Set API keys
cp .env.example .env
# edit .env and add ANTHROPIC_API_KEY and OPENAI_API_KEY

# 3. Pre-fetch market data (required before running the demo)
python3 -m pipeline.scripts.prefetch_markets

# 4. Run the evaluation
python3 app/main.py

# 5. Launch the dashboard
streamlit run app/dashboard.py

Run python3 -m pipeline.scripts.prefetch_markets to cache all demo markets. The script fetches 5 long-horizon and 5 speech markets and writes them to data/cache/. No live API calls are made during the demo.