Governed Proposal Steward

Agents that write metadata without review create trust debt. We draft a proposal, wait for human accept, then write back to DataHub.

Comment

Inspiration

AI agents that mutate catalog metadata without review create trust debt. Operators need a steward that proposes a change, waits for a human, then writes back to the DataHub graph so the next person (or agent) inherits context.

What it does

Governed Proposal Steward runs a fixed governance scenario:

  1. Before — demo dataset lacks gps.needs_governance_review.
  2. Draft — proposal with rationale + cited evidence.
  3. Human accept — apply is blocked until accept.
  4. After — Live Path writes the tag to local DataHub GMS and verifies it; Canned Path proves the same payload offline for judges without Docker.

How we built it

  • Python package (governed_proposal_steward) with CLI: health, run-scenario.
  • Dual-path: Live (ingestProposal + verify) and Canned (honesty labels).
  • Windows-safe seed script for the demo entity.
  • Offline pytest suite; Live smoke against DataHub quickstart (GMS :8080, UI :9002).

Challenges

  • Windows + DataHub CLI / Docker Desktop friction → documented Python 3.11 pin + seed script.
  • Keeping the demo honest: accept-before-write, dual-path labels, synthetic data only.
  • Filming a ≤3-minute story judges can evaluate without local setup.

Accomplishments

  • End-to-end Live Path: mode=live, verified_tag_present=true on the demo entity.
  • Canned path for offline review.
  • Public-facing Apache-2.0 package with README quickstart and examples/.
  • Optional MCP catalog adapter (same port; Accept stays in-process; offline-tested).
  • 147 offline tests; Live path verified against local DataHub quickstart.

What we learned

  • Load-bearing DataHub use = read context + action + write-back, not chat-over-catalog.
  • Human accept is the product, not a checkbox after the fact.
  • Dual-path keeps demos reproducible under judge time pressure.

What's next

  • Optional MCP Write-Back adapter is already in the package behind the same accept-then-apply port (GMS remains the default Live path and the film path).
  • More scenario fixtures (classification gaps, ownership, freshness).
  • Hardening Live verification lag and multi-tag merge for production catalogs. ```

AI / pre-existing code disclosure (adjust to truth)

Implementation and demo assets were developed with AI-assisted tooling under
human direction. Product code is original to this entry unless noted in the repo.
DataHub (Apache-2.0) is used as the catalog platform via local quickstart / GMS APIs.
No private production data. Synthetic demo entity only.

Track

Agents That Do Real Work

Built With

Share this project:

Updates