Assessments

E-23 Evidence Index

OSFI's model risk guideline and the AMF's AI guideline both take effect on May 1, 2027. The work between now and then is the same for every covered institution: prove, model by model, that the inventory exists, the risk rating is real, the review was independent, the cyber testing happened, the drift is watched, and the vendor contracts carry audit rights. This builder turns that work into a living file you can hand a board.

The job

A model risk file fails the same way every time: the model nobody listed, the rating nobody can defend, the review the builder ran, the test that never happened, the drift nobody watched, the vendor contract with no audit right. This builder walks seven evidence blocks for each model: the inventory (E-23 Pr. 2.1), the inherent risk rating (Pr. 2.2), the independent review (Pr. 3.4), the pre-deployment cyber assurance with B-13 s. 3.1.2 testing (Pr. 3.5), the drift monitoring (Pr. 3.6), the vendor audit rights (B-10 Principle 8, s. 2.3.3), and the AMF AI system entry where Quebec institutions are in scope.

You get a coverage index that scores each block complete, partial, or missing, a living-file JSON that follows the published schema v0.1.0, and a board-ready PDF. Every legal claim in the file carries a source id back to the guideline text.

What this is not

Not legal advice, not a regulatory filing, and not a certification of compliance. The file is an evidence format; the institution's accountable officers sign it and stand behind it. The builder also never decides whether you are compliant. It only shows which evidence is present and which is missing, so a reviewer can argue with the file instead of a blank page.

For the use case around an AI agent rather than the model itself, start with the Agent Access Scope Checker. For the program around the models, theAI Governance Starter Checklist and thePrivacy and AI Governance Maturity Assessment.

Start from a worked example, or clear it and build your own file.

A chat assistant answers policy questions on an insurer's site in Quebec. It never sees client records, but the AMF guideline covers any AI use, so it still needs an inventory entry, a risk rating, validation, and controls.

File name

Institution

Prepared by (role)

Prepared date

Model 1: Service chatbot (policy Q&A)

1. Model inventory

E-23 Principle 2.1: institutions should identify and track all models in use or recently decommissioned

Model name

Version

Description

Origin

Model owner

Model developer

Lifecycle status

Deployment date

2. Inherent risk rating

E-23 Principle 2.2: institutions should establish a model risk rating approach that assesses key dimensions of model risk

Risk rating

Rating date

Criteria used

Rationale

3. Independent validation

E-23 Principle 3.4: institutions should have a process to independently assess conceptual soundness and performance of models

Reviewer

Model approver

Is the reviewer independent from development?

Last review date

Next review date

Review outcome

Review scope

4. Pre-deployment cyber assurance

E-23 Principle 3.5 plus OSFI B-13 s. 3.1.2

Cybersecurity risk assessment date

Penetration test: last conducted

Penetration test: scope

Penetration test: provider

Penetration test: next scheduled

B-13 s. 3.1.2 names penetration testing and red teaming once and sets no fixed cadence. Your institution sets the triggers and minimum frequencies and defends them. Record the basis you would defend.

5. Continuous drift monitoring

E-23 Principle 3.6: institutions should have defined standards for model monitoring, and model decommission

Monitoring frequency

Last monitoring check

Drift detection method

Exceptions on record

6. Third-party audit rights

B-10 Principle 8 (s. 2.3.3)

Vendor name

Audit clause reference

Contract gives audit rights to the institution?

Vendor held to the institution's tech and cyber standards?

Last vendor audit

7. AMF AI system coverage

AMF Guideline for the Use of Artificial Intelligence (2026-PDG-0016), in force May 1, 2027

Does the AMF AI guideline apply to this institution?

Institution type

AI use that does not touch client records (e.g. a client chatbot)?

Validations performed

Controls in place

Evidence references

File-level: point each claim at the document or system where it lives.

The coverage index

Which evidence is present and which is missing, model by model. A block scores complete only when every required field carries a record. This is an index, not a compliance verdict.

Service chatbot (policy Q&A)

Model inventory (complete)OSFI-E23-2027
- All required fields carry evidence.
Inherent risk rating (complete)OSFI-E23-2027
- All required fields carry evidence.
Independent validation (complete)OSFI-E23-2027
- All required fields carry evidence.
Pre-deployment cyber assurance (complete)OSFI-E23-2027, OSFI-B13
- All required fields carry evidence.
Continuous drift monitoring (complete)OSFI-E23-2027
- All required fields carry evidence.
Third-party audit rights (complete)OSFI-B10
- All required fields carry evidence.
AMF AI system coverage (complete)AMF-2026-PDG-0016
- All required fields carry evidence.
- Flagged as an AI use that does not touch client records. The AMF guideline covers it anyway: any use of an AI system is in scope.

Export the file

The JSON follows the published living-file schema v0.1.0. The board summary prints as a one-page-per-model pack. Everything is generated in your browser. Nothing is stored or sent.

Read the board summary

E-23 / AMF AI GUIDELINE EVIDENCE FILE
Institution: [not named]
Prepared: 2026-11-07 by [role not named]
Both guidelines take effect on May 1, 2027.

Model: Service chatbot (policy Q&A)
  [COMPLETE] Model inventory (OSFI-E23-2027)
    - All required fields carry evidence.
  [COMPLETE] Inherent risk rating (OSFI-E23-2027)
    - All required fields carry evidence.
  [COMPLETE] Independent validation (OSFI-E23-2027)
    - All required fields carry evidence.
  [COMPLETE] Pre-deployment cyber assurance (OSFI-E23-2027, OSFI-B13)
    - All required fields carry evidence.
  [COMPLETE] Continuous drift monitoring (OSFI-E23-2027)
    - All required fields carry evidence.
  [COMPLETE] Third-party audit rights (OSFI-B10)
    - All required fields carry evidence.
  [COMPLETE] AMF AI system coverage (AMF-2026-PDG-0016)
    - All required fields carry evidence.
    - Flagged as an AI use that does not touch client records. The AMF guideline covers it anyway: any use of an AI system is in scope.

This living file is an evidence format. It is not legal advice, not a regulatory filing, and not a certification of compliance by movahedi.ca. The institution's accountable officers sign the file and stand behind its contents.

Scoping the engagement

The agent watches for change. It does not re-validate the model. Each cycle it checks whether anything in your evidence file moved: new models, new versions, drift flags, missed review dates. When it finds a change, it flags it and drafts the evidence update for a person to review. The independent judgment, the decision that the model is still sound, stays in the paid review. That is the part a machine cannot sign.

The file above is the watch layer: it shows what moved and what is missing. The paid review is the validate layer: independent judgment on whether each model is still sound, signed by a person. One does not replace the other.

Sources

OSFI-E23-2027: Guideline E-23: Model Risk Management (2027) (Office of the Superintendent of Financial Institutions, published 2025-09-11; Final; in force May 1, 2027). Primary source
OSFI-B13: Guideline B-13: Technology and Cyber Risk Management (Office of the Superintendent of Financial Institutions, published 2022-07-31; In force since January 1, 2024). Primary source
OSFI-B10: Guideline B-10: Third-Party Risk Management (Office of the Superintendent of Financial Institutions, published 2023-04-24; Revised guideline; in force May 1, 2024). Primary source
AMF-2026-PDG-0016: Guideline for the Use of Artificial Intelligence (Autorite des marches financiers (Quebec), published 2026-04-02; AMF decision March 30, 2026; published April 2, 2026; in force May 1, 2027). Primary source
BLG-2026-05: Quebec's AMF lays out its expectations for the use of AI (Borden Ladner Gervais, published 2026-05-12; Secondary commentary). Secondary commentary
NRF-2026-04: AMF's AI guideline is now official: What financial institutions need to know (Norton Rose Fulbright, published 2026-04; Secondary commentary). Secondary commentary
LEXOLOGY-2026-04: Insurance and Financial Services Distribution in Quebec: Regulatory Recap of the Past Six Months (Lexology, published 2026-04; Secondary commentary). Secondary commentary

How the coverage is scored

A block scores complete only when every required field carries a record. Partial means some fields are filled and the rest are blank. Missing means nothing is recorded. Not applicable covers the honest exceptions: an in-house model has no vendor arrangement to audit, a model still in development has no production drift to monitor, and a non-Quebec institution has no AMF entry to make. The index names the rule behind each score, including the one that bites hardest: E-23 Pr. 3.4 requires the review to be independent from development, so a reviewer who built the model cannot produce a complete block.

Questions

When do the two guidelines take effect?

May 1, 2027, for both. OSFI published the final Guideline E-23 on September 11, 2025 with an 18-month transition. The AMF adopted its Guideline for the Use of Artificial Intelligence by decision 2026-PDG-0016 dated March 30, 2026, published in AMF Bulletin vol. 23, no. 13 on April 2, 2026; it takes effect May 1, 2027.

Who has to comply?

E-23 applies to all federally regulated financial institutions, including foreign bank branches and foreign insurance company branches. It does not apply to federally regulated pension plans. The AMF guideline applies to authorized insurers, financial services cooperatives, authorized trust companies, and other authorized deposit institutions operating in Quebec.

Does the AMF guideline cover our client chatbot?

Yes. The AMF's final guideline applies to any use of an AI system, including uses that do not involve handling client records. A client chatbot that never sees a client file still needs an inventory entry, a risk rating, validation, and controls matched to that rating.

Does B-13 require penetration testing every three years?

No. B-13 s. 3.1.2 expects institutions to regularly perform intelligence-led tests and exercises such as penetration testing and red teaming, but the institution sets the triggers and minimum frequencies and defends them. The three-year cadence belongs to OSFI's intelligence-led cyber resilience testing framework for systemic institutions, not to B-13.

Where do my answers go?

Nowhere. The evidence file is built entirely in your browser. Nothing is stored, sent, or logged by this page.

Can the agent re-validate our models for us?

No. The agent watches for change: new models, new versions, drift flags, missed review dates. It does not re-validate the model. The independent judgment, the decision that a model is still sound, stays in the paid review. That is the part a machine cannot sign.

Is the French version of this tool ready?

Not yet. This page ships in English first; the French version is in preparation. Quebec coverage is claimed here with that stated openly.