# AI Model Lifecycle Stage Gates and Retirement Checklist

Stage gates from preview to retirement. Stage definitions with promotion criteria, a register of events that force re-assessment, a promotion gate checklist, and a retirement checklist.

**Who this is for:** The model governance owner defining what has to be true for a model to advance, and what pulls it back for review.

Source playbook: https://aigovernance.com/playbook/ai-model-lifecycle-governance

---

## Lifecycle stages and promotion criteria

_The stages your models move through and the evidence required to move between them._

### Template

| Stage | What it means | Promotion criteria to enter this stage |
|---|---|---|
| Preview | Available to a small internal group for evaluation | Registered; owner assigned; intended use documented |
| Approved for Testing | Wider internal use, non-production data paths | Preview eval complete; risk tier assigned; known limitations documented |
| Production (GA) | Live, affecting real decisions or users | Full evaluation (incl. bias where people are affected); human oversight design; Legal sign-off for High tier; rollback plan |
| Under Re-assessment | Still live, but flagged for review | A re-assessment trigger has fired |
| Deprecated | Discouraged, migration underway | Replacement identified; migration plan with dates |
| Retired | Off | Retirement checklist complete |

### Worked example

| Stage | Entered | Evidence at promotion |
|---|---|---|
| Preview | 2026-01-10 | Registered; owner D. Portillo; intended use documented |
| Approved for Testing | 2026-01-24 | Preview eval done; risk tier High; known limitations logged |
| Production (GA) | 2026-02-11 | Full eval incl. adverse-impact; oversight design; Legal memo; rollback to v2.3.1 |
| Under Re-assessment | 2026-08-30 | Base model changed VendorLM-3 to VendorLM-4 |
| Production (GA) | 2026-08-30 | Eval delta reviewed (ratio 0.85 to 0.88); cleared same day |

### Acceptance criteria

- The stages match how your organization actually ships models, not a generic template.
- Each promotion criterion is objective evidence, not a judgment call.
- Every production model has a current stage recorded in the registry.

---

## Re-assessment trigger register

_The events that move a live model to Under Re-assessment before it can continue._

### Template

| Trigger | Threshold | Who monitors | Action on fire |
|---|---|---|---|
| Model or base-model change | Any weights, prompt, or parameter change | MLOps | New version; re-run eval before continuing |
| Performance drift | <metric past its control limit> | Monitoring | Re-assess within X days |
| Bias metric shift | <fairness metric past threshold> | Monitoring | Immediate re-assess; consider pause |
| Use-case or population change | Any change to what it decides or who it affects | Business owner | Full re-classification |
| New or changed regulation | A regime now applies or its requirements changed | Legal | Re-assess against new requirements |
| Incident | Any Sev-1 or Sev-2 involving this model | Incident process | Re-assess as part of post-incident review |
| Time | No review in <N> months | AI Gov Lead | Scheduled re-assessment |

### Worked example

| Trigger | Threshold | Who monitors | Action on fire |
|---|---|---|---|
| Base-model change | VendorLM-3 to VendorLM-4 | MLOps | Fired 2026-08-30; eval re-run, cleared |
| Bias metric shift | Adverse-impact ratio below 0.80 | Monitoring | Not fired; current 0.88 |
| Time | 90 days since last review | AI Gov Lead | Next scheduled 2026-11-30 |

### Acceptance criteria

- Every trigger names who monitors for it and what happens when it fires.
- Thresholds are numeric where the trigger is a metric.
- A fired trigger is recorded against the model with a date and an outcome.

---

## Promotion gate checklist

_Run before a model advances a stage. Same checklist, applied at each transition._

### Template

| Check | Testing | Production | Evidence |
|---|---|---|---|
| Registered with all required fields | Yes | Yes | |
| Risk tier assigned | Yes | Yes | |
| Evaluation appropriate to the stage | Preview-level | Full, dated | |
| Bias evaluation (if people are affected) | Optional | Yes | |
| Human oversight design documented | Recommended | Yes for High | |
| Rollback plan | Recommended | Yes | |
| Legal sign-off | If triggered | Yes for High | |
| Owner accepts stage accountability | Yes | Yes | |

### Worked example

> Promotion: Approved for Testing to Production, fraud-scoring model.

| Check | Status | Evidence |
|---|---|---|
| Registered, risk tier assigned | Yes | High |
| Full evaluation, dated | Yes | 2026-09-01 |
| Bias evaluation (people affected) | Yes | no disparity above threshold |
| Human oversight design documented | Yes | analyst reviews all declines |
| Rollback plan | Yes | revert to rules engine |
| Legal sign-off | Yes | |
| Owner accepts stage accountability | Yes | Head of Risk Ops |
| Decision | Promoted | 2026-09-03 |

### Acceptance criteria

- The gate is run and recorded at every stage transition, not just the first.
- Each check has attached evidence, not a tick.
- A missing check blocks promotion absent a recorded exception.

---

## Retirement checklist

_Clean decommissioning: weights, logs, dependents, sign-off._

### Template

| Step | Owner | Done | Evidence |
|---|---|---|---|
| No production traffic or scheduled jobs remain | | Y / N | |
| Dependent systems migrated or sunset | | Y / N | |
| Weights deleted or archived per data policy | | Y / N | |
| Logs and evaluations archived to retention store | | Y / N | |
| Registry status set to Retired with date | | Y / N | |
| Inventory and documentation updated | | Y / N | |
| Owner sign-off | | Y / N | |

### Worked example

| Step | Owner | Done | Evidence |
|---|---|---|---|
| No traffic or jobs | MLOps | Y | 30-day zero-call log |
| Dependents migrated | MLOps | Y | 2 pipelines moved to the new model |
| Weights archived | MLOps | Y | Cold storage; prod deleted |
| Logs archived | AI Gov Lead | Y | 7-yr retention store |
| Registry status Retired | AI Gov Lead | Y | 2026-09-20 |
| Inventory updated | AI Gov Lead | Y | |
| Owner sign-off | Head of Risk Ops | Y | 2026-09-20 |

### Acceptance criteria

- Every step has an owner and evidence before retirement is called complete.
- Weight deletion or archival follows the organization's data retention policy.

---

## Governance controls this kit produces evidence for

- **MGV-001**: The stage definitions and promotion criteria are the preview and staged-release policy.
- **CHM-002**: The promotion gate checklist is the deployment gate process, applied at each transition.
- **CHM-005**: The retirement checklist is the model deprecation and decommissioning procedure.
- **MGV-002**: Stage entry criteria plus the gate checklist form the intake and approval workflow across the lifecycle.
- **MON-005**: The re-assessment trigger register defines the continuous-evaluation events that force a fresh review.
