Model Governance

Model governance records what a model is, where it came from, how it was evaluated, what use it was approved for, and which evidence supports its lifecycle decisions.

Unlike traditional ML registries that focus solely on operational deployment, the AI Governance and Evidence Infrastructure (AGEI) approaches models as highly regulated entities bound by strict lifecycle boundaries and evidentiary requirements.

Core Lifecycle Schema Objects

Model tracking relies on two primary relational constructs within the AGEI schema:

• ai_lifecycle_objects: Registers datasets, training runs, base models, fine-tuned model versions, evaluation bundles, and release artifacts. This ensures that receipts do not float without subject identity.
• ai_lifecycle_object_links: Complements receipt_links by preserving object-level lineage (e.g., mapping a specific model version backwards to its exact training dataset and evaluation benchmark).

Key Governance Stages

1. Dataset and Training-Run Lineage

Before a model exists, its inputs must clear a Provenance Gate. The system records dataset onboarding receipts and links them to subsequent training_run objects. This captures the exact code/config hash, the compute environment, and the output model to answer the fundamental question: "What data produced this behavior?"

2. Model Registration and Version Identity

Models are registered as immutable versions (e.g., claims_model_v4.3.2). They are never overwritten. A new training epoch or fine-tuning run produces a strictly distinct identity within the ai_lifecycle_objects table.

3. Evaluation and Benchmark Evidence

Promotion from staging to production requires passing a Validation Gate. The gate enforces policy rules (e.g., checking regression thresholds). A signed receipt is emitted that points to the raw validation matrices, test suites, and benchmark evidence stored in evidence_objects.

4. Risk Posture and Intended-Use Constraints

A model does not inherently have permission to operate universally. Models are registered alongside their risk classification and intended-use boundaries. Policy constraints restrict a model from being queried by certain applications or operating in high-risk jurisdictions without explicit authorization.

5. Release Approval and Artifact Release Records

Final promotion clears a Deployment Gate, committing a release approval receipt. For externally distributed models, artifact_release_records track the exact weights, signatures, and configuration packages released.

6. Retirement, Supersession, and Incident Linkage

When a model is deprecated, its lifecycle object is updated with a supersession link to its successor. If an anomaly or policy violation occurs, an Incident Gate is triggered. Generalized incident_links map the incident directly back to the specific model_version, its training data, and any affected downstream agents.