Skip to content

DISCLOSURES.md

Latest commit

 

History

History
45 lines (34 loc) · 2.25 KB

DISCLOSURES.md

File metadata and controls

45 lines (34 loc) · 2.25 KB
Raw

Disclosures

Creation period

Agent Evidence Graph was created for the Build with DataHub: The Agent Hackathon during the submission period. No pre-existing proprietary application code was incorporated.

Open-source software

The project uses open-source dependencies including DataHub Core and its Python SDK, DataHub Agent Context Kit, DuckDB, Pydantic, Typer, Streamlit, pytest, Ruff, mypy, and uv. Dependency versions are locked in uv.lock. The project source is licensed under Apache License 2.0; third-party components retain their own licenses.

AI assistance

AI coding assistants, including OpenAI Codex, were used for implementation, debugging, documentation, test generation, and release auditing. Human approval was retained for all real DataHub mutations and destructive reset/recovery steps.

The demo's English narration is AI-generated with Microsoft Edge online neural TTS via the open-source edge-tts client, using voice en-US-AndrewMultilingualNeural with scene-level rate and pitch adjustments. It does not clone or imitate a real person. The same disclosure appears in the demo package and on the final video frame. English subtitle timing comes from the synthesis stream's sentence-boundary events; Chinese translations are bound to the same timestamps and are not separately synthesized.

Synthetic and real evidence boundary

  • Finance business rows are deterministic synthetic demo data stored in DuckDB.
  • The eight semantic evaluation cases are fixed synthetic cases, not a production performance benchmark.
  • Offline adapter samples are fixtures and are labeled source_kind: fixture.
  • Listed real runs used a live local DataHub Core Quickstart instance.
  • Current real agent reads use official DataHub Agent Context Kit LangChain tools.
  • The demo-only runtime tag mutation and the human-approved governance writeback are separate real DataHub operations.
  • Screenshot and trace hashes identify reviewed artifacts; they are not digital signatures or a tamper-proof ledger.

Scope

This is a hackathon prototype for one agent, two versions, and two deterministic drift detectors. It is not presented as a production security control, a complete enterprise governance system, or a guarantee that every unsafe agent action will be detected.