Published

Model-change governance asks whether a particular version has been independently validated, authorised for a defined use, and released under the agreed conditions. This guide follows that evidence for Canada's federally regulated financial institutions, across traditional lending and actuarial models as well as AI and machine learning.
Canada's Office of the Superintendent of Financial Institutions (OSFI) has finalised the 2027 version of Guideline E-23, Model Risk Management, effective May 1, 2027. The lending-model example below connects a new cash-flow data source to its independent review, use restrictions, actual release, and later monitoring findings.
The companion OSFI document-agent guide covers runtime permissions, tool use, and actions on individual document cases. Here, the focus is the decision to adopt a changed model and the conditions for its continued use.
OSFI's E-23 backgrounder covers traditional models, including actuarial models, as well as AI and machine learning. The 2027 guideline applies to federally regulated financial institutions. Pension plans are excluded from this guideline's scope.
E-23 also addresses relevant foreign bank and insurance branches operating in Canada. Governance is risk-proportionate. Institutions identify models, while the enterprise inventory contains those with non-negligible inherent risk; an approved negligible-risk exemption still needs a tracked decision. These distinctions matter when scoping the evidence work.
The finalisation letter explains the transition from the earlier deposit-taking-institution guideline. Use the final 2027 text when planning that transition. The design below is a proposed evidence workflow, with institution-chosen identifiers and controls.
Give the change its own identity alongside the model ID and candidate version. Record the existing approved use, proposed difference, owner, affected population, and intended release. This lets a reviewer ask what changed without reconstructing the answer from a current inventory row and several moving folders.
Bind the review to identified document and artifact versions. Source code can stay in its controlled repository and datasets in their governed data environment; the change record can hold authorised references, selection details, and the evidence considered. Copying all underlying customer data into a general document package can create unnecessary access and retention problems.
| Record | Question the connection should answer |
|---|---|
| Change proposal | Which approved baseline is changing, for what purpose, and under whose ownership? |
| Candidate package | Which code, configuration, data selection, dependencies, and documents were submitted for review? |
| Review and response | Which version was challenged, what was found, and how were the findings addressed or accepted? |
| Use decision | Who authorised the candidate, with which use limits, conditions, and follow-up owner? |
| Release evidence | Does the deployed artifact match the approved candidate, and when did that use begin? |
| Monitoring and closure | What happened after release, what changed in the decision, and which version replaced or retired it? |
Preserve the selected package when new evidence arrives. A corrected report can be a later submission with a clear relationship to the earlier one. Keep the original assessment visible so a later reader can explain why a restriction or condition existed.
E-23 separates independent model review from the authority to approve use. Review triggers include model modifications, significant data changes, performance breaches, and scheduled reviews. Record the risk-based review scope and the separate use decision. A document-publication approval cannot substitute for that model assessment.
Capture the baseline, proposed change, intended use, selected evidence, and accountable owner.
Link reviewer findings and responses to the candidate, then record the authorised decision and conditions.
Connect the release, observations, exceptions, reassessment, and eventual replacement to that decision.
Make decision states useful to operators. “Review complete,” “approved for restricted use,” and “released” describe different events. The configured handoff should identify who can advance each state and what evidence they must consider. If the candidate changes during review, identify the new submission and its impact on the review scope.
The following lending-model scenario is fictional. Its identifiers, restricted population, and 90-day follow-up are institution-specific choices, not OSFI-prescribed thresholds or a general grace period.
Change CHG-00041 proposes version 3.2 of model MOD-00018, currently approved at 3.1. The owner identifies the cash-flow source, data-selection method, updated configuration, intended population, and affected dependencies. The proposal links to the method revision used for the assessment.
The reviewer receives candidate package PKG-00041 with identified tests, data-quality evidence, methodology, assumptions, and technical references. A later corrected test report becomes a new identified submission. The record shows which assessment considered it and how it changed the findings.
The reviewer identifies a limitation for applicants with limited history. After considering the review, the authorised approver permits restricted use with a manual-review condition and a 90-day reassessment. The owner and due date are visible operational work; the recommendation and approval remain distinct records.
The release owner compares the deployed version and configuration with the approved candidate. A mismatch returns to the change process. The release record links the accepted artifact, time, population, and conditions to the approval. A later operator can identify which decision governed a particular period of use.
A monitoring threshold is breached before the scheduled reassessment. The owner opens a linked investigation, and the responsible authorities consider the evidence and decide whether to restrict use further, remediate, re-review, or replace the model. Subsequent approval and retirement records preserve the earlier restricted-use history.
OSFI's finalisation letter says institutions should set criteria for material changes in self-learning models. It also provides no blanket grace period for validating third-party updates after deployment; an institution may define an exceptions policy for limited, specific uses. Record the actual authority and limits of any exception.
For a hosted model service, capture the provider version information available to you, relevant configuration, evaluation selection, and the change notice. Document uncertainty where a supplier does not expose a fixed artifact. The record should explain the assessment made with the available evidence and the resulting controls, without claiming reproducibility that the service cannot provide.
Include changes to data transformations, dependencies, configuration, and intended use in the change inventory where they affect the model's behaviour. Identify the candidate being assessed and record which changes the resulting approval covers.
Prepare an evidence path from a model change to the review, permitted use, actual release, and later observations. That connection makes the inventory useful to the people who must understand a past decision as well as the current status.
Regulatory sources reviewed: November 15, 2026. Regulatory statements refer to the linked final E-23 guideline, backgrounder, and finalisation letter. The evidence design and lending example are illustrative.
The open-source foundation — API-first, deployable into your own AWS account, and free to use. Right for architecture validation and early implementation.
Production-ready editions for departments and complex workflows. Start with a Proof-of-Value deployment or go straight to production.
For governance-heavy environments with residency, sovereignty, assurance, and multi-jurisdiction requirements. Talk to us about the right deployment model.