Governance: Lineage, Model Cards and Access

Answerable questions

Good governance means you can answer, for any prediction: which model version made it, trained on which data, with which code, evaluated how, approved by whom. Keep lineage (data to run to model version to deployment), model cards (purpose, data, metrics including per-group results, limitations, owner), access control on data, models and promotion rights, and audit logs. Regulated areas (credit, health, employment) may legally require explanations, documentation and human oversight.

Accountable, adaptable, reviewed

Governance makes models accountable; LLM systems add their own operational needs.

Lineage for one prediction

Each arrow should be queryable.

prediction p-88231 (2026-09-30 14:02)
  <- model cancer-clf v7 (alias champion since 2026-09-12, approved by risk-ml lead)
  <- run 4f2a... (git 9c1e2d, params logged, data fingerprint 04cf6d426dd1)
  <- dataset snapshot 2026-09-01 (validated, 412,331 rows)
  <- evaluation report v7 vs v6 (overall + slices, gate passed)

Restrict who can promote

Separate who can train models from who can move the production alias, and log every promotion.

Quick check: What does lineage let you answer?

• Which data, code and approvals produced the model behind a prediction
• Which colour the dashboard uses
• How many GPUs exist
• The CEO's schedule