Skip to content

Model and AI governance infrastructure for OSFI FRFI

   days left

Can you prove that your models and AI are governed?

MINERVA Studio: We make you OSFI E‑23 compliant.

Contact us
FilmMINERVA — E-23 readiness, walked (4 min)
Open the film in a tabEsc closes
app.minerva.studio › Map

01 · System of recordsEvery model, owner, policy, vendor and decision sits on one record, with each change hash-chained to the one before.

02 · InventoryEvery model is held in E‑23's own fields: owner, developer, reviewer, approver, risk rating and next review date.

03 · Policy as codeThe guideline's requirements are written as a rule pack, and the record is checked against them on every change.

04 · AssessmentEach model is risk-rated on E‑23's factors by a named person, and the rating sets its validation, monitoring and approval requirements.

05 · ValidationTests run in your own environment on a plan set by the risk rating, and every result is recorded against the model.

06 · System of actionsNo model is approved with a segregation-of-duties conflict, missing evidence or an open high-severity finding, and only a named person approves.

07 · Evidence packetOne package holds every record, time-stamp receipt and signed checkpoint, and your examiner verifies it offline with a single command.

08 · Audit portal / trailEvery change is attributed and time-stamped by independent authorities, and examiners get time-limited, read-only access.

09 · Enterprise supportRun it hosted in Canada, on a dedicated instance or on your own servers, with single sign-on and an open API to the same record.

10 · Rule packsRule packs for OSFI E‑23, SR 26‑2, PRA SS1/23, the NIST AI RMF and the EU AI Act set the fields and approval rules the record must meet.

WHERE IT COMES FROMA spreadsheet importinventory.xlsxA registry pullMLflow · Archer · OpenPagesA person's editevery change attributedTHE RECORDHASH-CHAINEDMODELIFRS 9 ECL overlaydrawn by Yusuf Bellob81e04c2⑆69ad53e5PERSONLena Fischerline 2 · model reviewer5d3a91f0⑆b81e04c2POLICYModel risk policyversion 3c4dbc2c0⑆5d3a91f0VENDORFeaturespace Ltdassessed this year7e9aa8c5⑆c4dbc2c0DECISIONApproved for useRajiv Menon · senior management4a0f5797⑆7e9aa8c5ANYONE CAN RECOMPUTEsha256(the row, canonical)= 4a0f5797…each row holds the prior row's hashIMPORTinventory.xlsx13 rows · 10 E‑23 columnsINVENTORYMINERVA BANK · DEMOMODELDEVELOPERREVIEWERMarket risk VaRMNB-VAR-009Priya SharmaLena FischerIFRS 9 ECL overlayMNB-ECL-005Yusuf BelloYusuf BelloCustomer assistantMNB-GEN-013 · AIMicrosoftMiguel SantosAML monitoringMNB-AML-011NICE ActimizeMiguel SantosCard fraudMNB-FRD-012FeaturespaceLena FischerCONFLICT5 OF 13 MODELS · E‑23 FIELDSE‑23 FIELDSOwnerDeveloperReviewerApproverRatingNext reviewTHE GUIDELINEOSFI E‑23Independent review before approvalValidation scaled to the ratingApproval by the named authorityNo approval over an open findingin force 1 May 2027POLICY AS CODEONE PACK · ONE ENGINEpack: osfi-e23move: approve requires_separation: [developer, reviewer] requires_group: [validation] requires_authority: true blocked_by: finding ≥ high, openband: high review_every_days: 365 monitor_every_days: 31 approve_role: senior_managementRead on every change, over 13 modelswhat a door refuses, it says why; what is owed prints in the gap reportTHE RULE IS DATA · THE ENGINE ALLOWS OR REFUSESCHECKED ON THE RECORDAPPROVEIFRS 9 ECL overlayreviewer ≠ developerYusuf Bello holds bothvalidation evidencefour suites owedthe named authorityRajiv Menon · senior managementno open high findingstability · openAPPROVE · SHUT · 3 REASONS PRINTEDRATED BYHZHelen Zhaohead of model risk · line 2RATINGIFRS 9 ECL OVERLAYPotential financial impactHIGHComplexity, and level of autonomyMEDIUMBusiness use or purposeHIGHReliability of the data it takes inMEDIUMImpact on customersLOWRegulatory riskHIGHBAND · A JUDGMENT OVER THE FACTORSHighre-rate by 14 Oct 202710 FACTORS · E‑23 PRINCIPLE 2.2WHAT HIGH ASKSVALIDATIONevery 12 monthsMONITORINGevery 31 daysRE-RATINGevery 12 monthsAPPROVALsenior managementTEST SUITESnineYOUR NOTEBOOKMINERVA-MRMrun = mn.run(model["id"], suite="stability")run.add(suites.stability(X_train, X_recent))run.record(open_findings=True)VALIDATION PLANHIGH · NINE SUITESPerformancePASSStabilityMIXED · FINDINGData qualityPASSCalibrationPASSBacktestingPASSSensitivityOWEDBenchmarkOWEDSegmentsOWEDFairnessOWEDLFLena Fischer signs off, as the reviewerthe developer cannot sign offRUN IN YOUR NOTEBOOK · RECORDED HEREFINDING · OPENED BY THE RUNStability: 4 of 16 failedHIGH · OPENblocks approval until it is closeda later pass is stamped as a retestTHE SOFTWARE PROPOSESBABuilt-in analystmay propose; may not decide or signREVIEW & APPROVALIFRS 9 ECL OVERLAYApprove for useMNB-ECL-005 · HIGH · SENIOR MANAGEMENT DECIDESOwner and approver namedValidation evidence on fileLimitations and uses recordedConflict: Yusuf Bellodeveloper and reviewerIndependent reviewLena FischerBLOCKED · ASSIGN AN INDEPENDENT REVIEWERCLEARED · SENIOR MANAGEMENT DECIDESBABuilt-in analyst proposed itAPPROVEAPPROVEAPPROVEDR. MENON · 14 OCT 2026THE SOFTWARE PROPOSES · A PERSON DECIDESA PERSON DECIDESRMRajiv Menonsenior management · not the proposerTHE EVIDENCE DESKEXPORT THE PACKETthe export is itself a rowTHE PACKETEVIDENCE-PACKET.ZIPREADME.mdledger/ledger.jsonregisters/dossiers/receipts/ · 4checkpoints/verify.mjs$ node verify.mjsok: 3060 events, head 4a0f5797…ok: 127 files match SHA256SUMSok: 4 anchor receipts name a ledger rowEVERY RECORD, RECEIPT AND CHECKPOINTTHE EXAMINER'S MACHINENode.js and a shellnone of our softwarechecked offline, without usEXAMINER ACCESSEXOSFI examinerreads every row; proposes nothingREAD-ONLY · UNTIL 31 MAR 2027THE TRAILNEWEST FIRST3,060RMRajiv Menonapproved for use4a0f5797⑆7e9aa8c53,059BABuilt-in analystproposed: approve for use7e9aa8c5⑆c4dbc2c03,058HZHelen Zhaonamed Lena Fischer reviewerc4dbc2c0⑆69ad53e5TIME-STAMPEDTIME-STAMPEDTIME-STAMPEDfreetsa.org · 2026-10-14 18:00:58ZITS OWN CLOCKCheckpoint: the day's head, signedED25519WHO · WHAT · WHEN · TIME-STAMPTIME-STAMP AUTHORITIESfreetsa.orgtimestamp.sectigo.comWHERE IT RUNSHosted · AWS Canada (Central)Montréal · the records stay in CanadaA dedicated instanceone institution, the same regionYour own serverson-prem, your registry exportsOPEN APIV1GET /api/v1/nodes/{id}/runs200 OK{ "suite": "stability", "status": "mixed", "counts": { "pass": 12, "fail": 4, … }, "by_name": "Lena Fischer", "hash": "de7671b4…"}THE SAME RECORD, BEHIND THE SAME GATESSIGN-INSingle sign-onOIDC · SAML · SCIMthrough your identity providera second admin approves sign-in changesWHAT A PACK CARRIESFIELDSwhat the record must holdAPPROVALSwhen a model may moveGATESwhat an overlay addsRULE PACKSSEVENOSFI E‑23model risk management · CanadaAI riskNIST AI RMF · ISO/IEC 42001SR 26‑2model risk management · USPRA SS1/23model risk management · UKEU AI Actthe high-risk tierOSFI B‑10third-party riskInformation assetsthe asset registerRULE PACKS ARE DATA · ONE ENGINEAPPLIED TO THE RECORD13 modelsgaps printedthe gap report, by the pack's own words

AI, security and GRC

Trusted AIAccuracyAccountabilityControlGatewayPrompt defenceAccess controlGuardrailsValidationModel testingIndependent reviewMonitoringAuditControl monitoringAccess reviewsAttestationVendorsHiddenLayerNoma SecurityStraikerVendorsValidMindYieldsSASVendorsVantaDrataSecureframeAIImplementationOpenAIAnthropicGoogle DeepMindFrontier modelsCopilotsAgentsSecurityProtectionPalo Alto NetworksFortinetCiscoThreat detectionZero trustIdentityGRCInventoryMINERVAMINERVA providesCompliance servicesGRC platformTo make you E‑23 compliantMINERVA SOFTWARE INCAI assurance09 Oct 2026
Model inventoryEvidence packetsAudit readinessforNIST AI RMFOSFI E‑23SR 26‑2
AI Implementation
OpenAI
Anthropic
Google DeepMind
Security Protection
Palo Alto Networks
Fortinet
Cisco
GRC Inventory
MINERVA
AISecurityGatewayValidationAuditTrusted AIGRCMINERVA

  1. AI Implementation

    • OpenAI
    • Anthropic
    • Google DeepMind

    Frontier modelsCopilotsAgents

  2. Security Protection

    • Palo Alto Networks
    • Fortinet
    • Cisco

    Threat detectionZero trustIdentity

  3. Gateway

    Prompt defenceAccess controlGuardrails

    Vendors

    • HiddenLayer
    • Noma Security
    • Straiker

  4. Validation

    Model testingIndependent reviewMonitoring

    Vendors

    • ValidMind
    • Yields
    • SAS

  5. Audit

    Control monitoringAccess reviewsAttestation

    Vendors

    • Vanta
    • Drata
    • Secureframe

  6. Trusted AI

    AccuracyAccountabilityControl

  7. GRC Inventory

    MINERVA

    Model inventoryEvidence packetsAudit readiness

    forNIST AI RMFOSFI E‑23SR 26‑2

    MINERVA provides

    Compliance servicesGRC platform

    To make you E‑23 compliant

Two plans.

One goal: Govern, Map, Measure and Manage.

What we deliver.

01 / 07
app.minerva.studio › Home
Home on the demo bank, the whole window: the programme down the left rail; where the bank stands, read from the record with its last outside witness; the gap report, the board report and the evidence packet one press away; the seven steps as one band with their live counts; what waits on the owner, a breach and a blocked approval first; and what the record owes today.

Programme Dashboard

Your E‑23 programme in seven steps, with live counts from the record and your open actions listed first.

app.minerva.studio › Inventory
The inventory: thirteen models, twelve carried, one identified only, one retired and retained and 146 fields owed; Import spreadsheet, Columns, Gap report and Add a model; the saved views; and the first models with their rating, where each stands, owner, model reviewer and next review date, with a review owed and a reviewer who is also the developer flagged in red.

Model Inventory

Every model with its owner, risk rating, reviewer, approver, next review date and monitoring status in E‑23's own fields.

app.minerva.studio › Review & approval
A decision card on the review page: approval for use of the IFRS 9 ECL macro overlay, asked by Anita Desai, with each condition still open beside its fix (the approver and the approved uses not yet recorded, one person named as both developer and reviewer, validation evidence owed, senior management's authority required), the conditions already met, and Reject with a reason as the only button.

Approval Controls

Approval stays blocked until every condition is met, and the card lists each blocker with what resolves it.

app.minerva.studio › Review & approval
Monitoring: one breach in production, twelve watches on eight models and no model in production without one; Run the sweep; and the watches with breaches first, the Card behavioural score's PSI at 0.29 against its line of 0.25 and the customer assistant's hallucination rate at 0.04 against 0.02, then three holding.

Monitoring and Escalation

When a monitored metric crosses its threshold, the system proposes a return to review and a named person decides.

app.minerva.studio › Evidence › The trail
The trail filtered to the Card behavioural score: the head strip with the rows on the record, the head's sequence number, the chain's verdict, 100 receipts from two authorities, the last time-stamp and the checkpoints; and two rows, the built-in analyst proposing to send the model back for review and the owner rejecting it with her reason, each with its hash seam, the time-stamp authority's own clock and the checkpoint that holds it.

Audit Trail

Every change is attributed, hash-chained to the previous one and time-stamped by independent authorities.

app.minerva.studio › Evidence › The examiner
The evidence packet's preview, open over the evidence desk: the files down the left, the README, the manifest, SHA256SUMS and verify.mjs first, then the registers; the README open, naming the bank, when and by whom it was exported and the ledger head with its hash; and Download for the zip.

Evidence Packet

Every record, receipt and signed checkpoint in one file, which your examiner verifies offline without access to our systems.

app.minerva.studio › Map
The demo bank's map at working zoom: the Model Risk Management team's card, led by Helen Zhao, above the Model Risk Management Policy (E‑23) at version 4, the model risk rating methodology at version 2 and the monthly control CTL-MRM-02, ongoing monitoring against recorded lines, with the strings between them labelled owns and governs.

TCoE Blueprint

Your model risk function on one map: teams, the E‑23 policy, the rating methodology, controls and how they connect.

Contact us

Guides In plain language

Contact us for exclusive deals.

FormContact
Which way

A person replies within one working day