# Synthetic Name Screening Tool-Use Dataset v1

This dataset provides 120 multilingual, synthetic trajectories for teaching or evaluating an AI agent that calls the public financial-crime.ai Name Screening tools.

## Scope

- Every subject and prompt is synthetic.
- No user query, live report, provider response, personal data, sanctions record, PEP record, or adverse-media result is included.
- Each case requires `screen_name`, then `get_screening_report` until `final_report_ready=true`.
- A pending result is never a final report.
- `no_reliable_evidence` must not be rewritten as clearance.
- Ten deterministic trajectory classes cover Unicode width, punctuation, entity type, missing optional fields, repeated polling, transient errors, and release changes.
- Terminal failure trajectories must not emit a stale or partial report.

## Files

- `cases.jsonl`: one JSON object per synthetic training or evaluation case.
- `manifest.json`: contract, integrity hashes, usage boundaries, and licensing metadata.
- `LICENSE-DATA.txt`: CC BY 4.0 data license notice.

Use the current function definitions from `/public/agents/function-toolkit.json`. Live calls remain subject to public rate limits and abuse controls.

Regenerate the JSONL and manifest hashes with `python ops/ai_distribution/generate_name_screening_tool_use_dataset.py`.