SCENE 04 — GOVERNANCE FLOW

OpenMetadata — Governance Plane

Mapping AI interpretability artifacts into a familiar metadata control plane

🏗️
Topology Mapping
Model (Service) Database (Model) Schema (Transformer) Layer_1 Layer_2 Layer_N Head_1 Head_2 Head_3 DEFECTIVE QUARANTINED create entities
DEFECTIVE QUARANTINED AUDITED VERIFIED
💡 Why Column-Level Tags?
Each attention head maps to a column. Enables fine-grained governance, existing data team workflows, and familiar query patterns.
🔄
Governance Flow
1
Bootstrap Topology
Backend creates synthetic OpenMetadata entities: model → schemas → layers → heads
2
Trace Ingestion
AttentionTrace objects ingested as lineage edges: prompt → layer → head → token
3
Operator Tags Head
User marks head as DEFECTIVE or QUARANTINED via OpenMetadata UI or REST API
4
Runtime Enforcement
Backend polls/receives webhook → calls set_masked_heads() → HeadMaskStore updates → subsequent generations masked
🔧 Engineering Robustness
Tagging includes retry/backoff logic. Prefers column-level tags, falls back to table-level when needed. Webhook handler validates secrets before processing.
← Demo Video Thank You → Tech Stack Intro
Topology Mapping — Full View
Model (Service) Database (Model) Schema (Transformer) Layer_1 Layer_2 Layer_N Head_1 Head_2 Head_3 DEFECTIVE QUARANTINED create entities