UBC — AI Solutions

Audit — Dynamic Documentation Platform

Living documentation that stays in sync with engineering, cloud, AI, and compliance activity. Documents update themselves as the organization changes.

Back to library
SOP
Fresh

SOP-AIDLC-07 — Deployment, Canary & Model Registry

Deploys models via the registry with canary and staged rollouts.

Owner: AI Platform DirectorApprover: Engineering DirectorVersion: 2.2Updated: 12/7/1969

1. Purpose

Deploy AI systems safely with reproducible provenance.

2. Scope

All AI deployments to staging or production.

3. Definitions

  • Model registry — The system of record for models, versions, and metadata.
  • Canary — A limited-audience deployment used to validate a model or prompt change.

4. Roles & Responsibilities (RACI)

ActivityRACICadence
Register model versionAI EngineerAI Platform DirectorMLOpsTeamPer version
Approve promotionAI Governance CouncilAI Platform DirectorGRCTeamPer promotion
Execute canaryMLOpsAI Platform DirectorAI EngineerTeamPer deploy

5. Procedure

  1. Register every candidate model version with metadata (dataset, eval, code hash).
  2. Promote only versions with signed evaluation and required approvals.
  3. Deploy as a canary to a defined slice (e.g., 5%) with monitoring.
  4. Roll forward or roll back based on canary metrics.

6. Inputs & Outputs

Inputs

  • Signed Evaluation Report
  • Model Card

Outputs

  • Registered model version
  • Deployment record

7. Controls & Metrics

MetricTarget
Deployments with canary100% for Tier 1/2
Rollbacks within 30 min of red-flag≥ 95%

8. Exceptions & Escalation

  • Prompt-only changes may deploy without canary if evaluation delta is < 1%.

9. Records & Retention

RecordRetention
Registry entriesLife of model + 5 years
Deployment records5 years
  • SOP-SDLC-07
  • SOP-AIDLC-05

11. References

  • ISO/IEC 42001
  • MLOps SIG

12. Revision History

See the Versions tab for the full change history maintained by the Auto-Doc Engine.


Document code: SOP-AIDLC-07 · Aligned to MLOps / ISO/IEC 42001. Controlled document — reproduction outside the UBC QMS requires the Quality Manager's approval.

Suggested by AI
No suggestions right now.
Frameworks
AI Governance
CMMI-DEV
SOC 2