# AI Model Registry Template and Entry Checklist The registry as a governed source of truth, not a spreadsheet nobody updates. A field schema, a pre-deployment entry gate, a versioning policy that says what triggers a new record, and a retirement process. **Who this is for:** The MLOps or governance owner standing up a model registry and wiring it to approval and monitoring workflows. Source playbook: https://aigovernance.com/playbook/ai-model-registry --- ## Model registry data schema _The fields every entry must carry. Required fields gate deployment; optional fields add context._ ### Template | Field | Required | Description | Example | |---|---|---|---| | model_id | Yes | Stable identifier | mdl-fraud-scoring | | version | Yes | Semantic version of this record | 3.2.0 | | owner | Yes | Named accountable individual | | | use_case | Yes | What decision or output it produces | | | lifecycle_status | Yes | Preview / Testing / Production / Deprecated / Retired | | | risk_tier | Yes | From the classification method | | | training_data | Yes | Sources, cut-off, lawful basis | | | applicable_regulations | Yes | Named regimes | | | eval_summary | Yes | Latest performance, and bias where people are affected, with dates | | | last_review | Yes | Date of last governance review | | | upstream_dependencies | No | Base models, vendors, datasets | | | deployment_targets | No | Where it runs | | ### Worked example | Field | Value | |---|---| | model_id | mdl-resume-screen | | version | 2.4.0 | | owner | D. Portillo, Head of Talent | | use_case | Ranks applicants 1-5 for recruiter triage | | lifecycle_status | Production | | risk_tier | High | | training_data | 5 yr hiring outcomes (internal), base model VendorLM-4; lawful basis legitimate interest with notice | | applicable_regulations | EU AI Act Annex III, NYC LL144, GDPR | | eval_summary | Accuracy 0.81 (2026-08); adverse-impact ratio 0.88 (2026-08), above the 0.80 line | | last_review | 2026-08-30 | | upstream_dependencies | VendorLM-4 base model; Greenhouse ATS | ### Acceptance criteria - Every required field is defined with a data type and an allowed-values list where relevant. - The schema is the same one the deployment gate checks against. - Regulation and eval fields carry dates, so staleness is visible. --- ## Registry entry pre-deployment checklist _The hard gate. No model reaches production without a complete entry._ ### Template | # | Check | Status | Blocker if not met | |---|---|---|---| | 1 | All required schema fields populated | Y / N | Yes | | 2 | Risk tier assigned via the classification method | Y / N | Yes | | 3 | Latest evaluation attached, dated within 90 days | Y / N | Yes | | 4 | Human oversight design documented (High tier) | Y / N | Yes for High | | 5 | Applicable regulations confirmed with Legal | Y / N | Yes | | 6 | Owner has accepted accountability in writing | Y / N | Yes | | 7 | Rollback plan recorded | Y / N | Yes | ### Worked example | # | Check | Status | Note | |---|---|---|---| | 1 | Required fields populated | Y | | | 2 | Risk tier assigned | Y | High | | 3 | Evaluation within 90 days | Y | 2026-08-30 | | 4 | Human oversight design documented | Partial | Override exists; no gate on low scores. Blocker. | | 5 | Regulations confirmed with Legal | Y | Memo 2026-03 | | 6 | Owner accepted in writing | Y | | | 7 | Rollback plan | Y | Revert to v2.3.1, 15 min | | **Gate decision** | | **Blocked** | Close item 4 before deployment | ### Acceptance criteria - The checklist is enforced in the deployment pipeline, not run on the honor system. - A failed check blocks deployment unless a named approver records a time-boxed exception. --- ## Model versioning policy _Says which changes create a new registry version and which of those need a fresh governance review._ ### Template > One page. Attach to the schema. **A new version record is created when any of these change:** - Model weights (retrain, fine-tune, base-model swap) - The prompt or system instructions, for prompt-based systems - Inference parameters that affect output (temperature, thresholds, retrieval config) - The training or retrieval data set - The intended use case or the population affected **A new version also triggers a governance review when:** - The risk tier could change - The change affects a High-tier system - Evaluation shows a material shift in performance or bias - A new regulation now applies **Every version record carries:** version number, date, author, changelog, evaluation delta, reviewer (if triggered). ### Worked example **Change: swapped base model from VendorLM-3 to VendorLM-4 on the resume screener.** - New version: 2.4.0 (weights changed). - Governance review triggered: yes (High-tier system; eval delta). - Changelog: base model upgrade; re-ran adverse-impact eval, ratio moved 0.85 to 0.88. - Reviewer: R. Nkemelu, 2026-08-30. Outcome: approved, monitoring cadence unchanged. ### Acceptance criteria - The list of change types that create a version is specific enough that two engineers would agree. - Every production model in the registry has at least one version record with a changelog. - Review-triggering changes have a linked reviewer and outcome. --- ## Model retirement process _Decommissioning steps so a retired model does not linger in code, storage, or documentation._ ### Template | Step | Owner | Done | Evidence | |---|---|---|---| | Confirm no production traffic and no scheduled jobs call the model | | Y / N | | | Migrate or sunset dependent systems | | Y / N | | | Delete or archive model weights per the data policy | | Y / N | | | Archive documentation, evaluations, and version history | | Y / N | | | Set registry lifecycle_status to Retired with a date | | Y / N | | | Notify stakeholders and update the AI inventory | | Y / N | | | Owner sign-off | | Y / N | | ### Worked example | Step | Owner | Done | Evidence | |---|---|---|---| | Confirm no traffic or jobs | MLOps | Y | 30-day traffic log, zero calls | | Migrate dependents | MLOps | Y | Legacy chatbot moved to support copilot | | Delete or archive weights | MLOps | Y | Weights archived to cold storage, prod copy deleted | | Archive documentation | AI Gov Lead | Y | Moved to governance archive, 7-yr retention | | Set status to Retired | AI Gov Lead | Y | 2026-09-15 | | Notify and update inventory | AI Gov Lead | Y | Inventory row marked retired | | Owner sign-off | Head of Support | Y | 2026-09-15 | ### Acceptance criteria - Retirement is not complete until every step has evidence and an owner sign-off. - The registry and the AI inventory both show the model as retired with a date. --- ## Governance controls this kit produces evidence for - **CHM-001**: The schema and versioning policy are the model version control standard, with a version record per change. - **CHM-004**: Each version record's changelog and evaluation delta are the model change documentation. - **CHM-005**: The retirement process is the model deprecation procedure with evidence per step. - **MGV-002**: The pre-deployment entry checklist is the intake and approval gate. - **SCT-009**: The populated registry serves as the algorithm register for production models.