docs/schema.md

# Schema

Each JSONL record in `metadata/` contains the following fields:

- `id`: unique sample identifier.
- `split`: dataset split (`train`, `val`, `test`).
- `domain`: one of `finance`, `medical`, or `iiot`.
- `video`: relative path to the MP4 file.
- `audio`: relative path to the WAV file.
- `frame_preview`: relative path to a representative PNG frame.
- `duration_sec`: clip duration in seconds.
- `fps`: video frame rate.
- `sample_rate`: audio sampling rate.
- `task_type`: benchmark task family.
- `question`: generated question.
- `ground_truth`: deterministic answer.
- `has_conflict`: whether a counterfactual cross-modal conflict was injected.
- `source_mode`: substrate source (`synthetic` or `yfinance` for finance in the full generator).
- `seed`: random seed used by procedural generation.
- `metadata_file`: relative pointer to the split metadata file.
- `qa_file`: relative pointer to the split QA file.
- `algorithmic_metadata`: domain-specific gradients and trigger values used by the state machine.

The `qa/` JSONL files mirror the QA-specific subset of this information.