# Replay Dataset for Deterministic Extraction Testing

Archived document fixtures used to verify that the extraction pipeline
produces consistent, schema-valid results across code changes.

Each fixture is a JSON file containing:
- `document_id`: stable identifier for the fixture
- `document_type`: article, filing, transcript, or press_release
- `document_text`: normalized text as it would arrive from the parser
- `known_tickers`: ticker hints passed to the extraction prompt
- `expected_extraction`: the expected extraction result (schema-valid)
- `metadata`: fixture provenance info (created_at, description, schema_version)

The replay runner (`tests/test_replay_extraction.py`) loads these fixtures,
validates the expected outputs against the current extraction schema, and
optionally runs them through a live Ollama instance for end-to-end checks.