Free tool Tier a model against OSFI E-23 in about five minutes — seven risk dimensions, the controls your tier requires, and a report you can print. Open it →
osfi-gate

Compliance as code · For Canadian banks, insurers and credit unions · OSFI B-13 · B-10 · E-21 · E-23

  • OSFI B-13
  • OSFI B-10
  • OSFI E-21
  • OSFI E-23 · now in ServiceNow

The audit evidence is written before anyone asks for it

osfi-gate runs in your build pipeline on every change. It maps what your scanners find to controls anchored to OSFI B-13, B-10, E-21 and E-23, applies the policy your risk team owns in Git, and files a signed record of the decision. Nobody has to remember to do anything.

See a year of evidence Ask a question

EVIDENCE INSPECTOR / PUBLISHED BUNDLESIGNED
repository
northwind/adjudication-service
commit
1b9bf83f9cfd
policy
f210e70e221d at 4f437fb50c6a
model
MRM-0204 · high materiality
servicenow
TASK0020860 · Awaiting risk approval
decision
FAIL / recorded
digest
sha256:6f56aa9795e4…

$ osfi-gate verify ./evidence ✓

A bundle published on this site. Download it and check every line.

Four steps, on every change

Your controls run the way your infrastructure already does: declared as code, reviewed in a pull request, versioned. Nothing runs on a schedule somebody has to remember, and nothing waits on a person filling in a form afterwards.

  1. It runs in your pipeline. On every pull request and every push to a protected branch, next to the scanners you already have: Semgrep, Trivy, Gitleaks, Checkov. It adds its own checks for dependency pinning, licences, and the places your code calls an AI model.
  2. Your policy decides, not ours. Each finding is matched to a control identifier. The rules that decide what blocks a build and what only warns are code, in a repository your second line owns — reviewed in a pull request and versioned like anything else. Not a vendor's opinion of your risk appetite.
  3. The evidence writes itself. Every run records what the scanners found, which version of the policy applied, what was decided, and everything needed to reach that decision again. It is hashed, signed, and written where it cannot be edited afterwards, including by us.
  4. Anyone can check it, later, on their own. One command re-hashes the files, checks the signature and the chain, and re-applies the recorded policy to the recorded inputs to confirm it still reaches the same result. Offline, with nothing of ours running.
osfi-gate verify --evidence-dir ./evidence --repo owner/name \
    --policy-dir ./policy --pubkey ./key.pub

The console reads that evidence. It is a viewer, not the source of truth. More on compliance as code →

New · E-23 model risk bridge

Model risk, filed where your risk team already works

The gate finds model calls in code. The bridge puts them into ServiceNow: an inventory record linked to the owning configuration item, a risk assessment somebody has to approve, and E-23 drafts written from what is actually in the repository. The pull request stays blocked until that assessment is approved.

GitHub Action

Detects on the pull request. The same detection that already registers AI usage against MR-02 — not a second scanner with its own opinion about what is in your code.

MCP server

Writes to ServiceNow. An inventory record as a configuration item, related to the application that calls it, and a risk assessment keyed so that running twice leaves one of each.

Claude agent

Drafts the E-23 artifacts. Purpose, data lineage and a monitoring plan, each citing the file it came from, so a reviewer opens the assessment and finds the reading already done.

Required check

Merge waits on risk. MR-08 fails the gate until the assessment is approved. An unreachable system is not an approval, and neither is a missing one.

Three things it will not do: compute the materiality tier, which the policy recomputes from the seven ratings so a record cannot understate its own; approve its own assessment, which is the second line's act in their own system; or invent a configuration item that no discovery source owns. Each refusal is in the code, not in a prompt.

What it actually puts in ServiceNow

These are records the bridge wrote, on a ServiceNow developer instance, on 27 September 2026. Invented data The model is Northwind Trust's — the fictional bank behind the demo estate — so what these show is that the bridge works, not that a bank is running it.

A ServiceNow configuration item form titled OSFI AI Model, named northwind/adjudication-service:ai/score.py, with Inventory ID MRM-0204, Repository northwind/adjudication-service, Call site ai/score.py, Provider self-hosted, Materiality (from gate) high, Evidence bundle 000065-1b9bf83f9cfd and Policy commit f636df7bfddf.
The inventory record, as a configuration item. The tier reads from gate because the policy computed it from the seven E-23 ratings; the bridge copies it and is not allowed to decide it. It names the evidence: the bundle that recorded the decision and the policy commit that applied, so a reviewer can get from this record back to the run and replay it. The bundle field arrives in a second call — a bundle id only exists once the bundle is written, which is after the decision the record was opened for.
A ServiceNow CI Relationships list showing one row: parent Adjudication Service, type Uses::Used by, child northwind/adjudication-service:ai/score.py.
The model related to the application that calls it. The application CI already existed; the bridge refuses to create one, because a CMDB filling up with records no discovery source owns is how a CMDB stops being believed.
A ServiceNow task TASK0020860, state Open, short description OSFI E-23 risk assessment: MRM-0204, with a description listing repository, call site, inventory ID, provider, model and materiality, and an offline replay command.
The assessment somebody has to approve. It carries the command to replay the decision offline, so a reviewer can check the gate rather than take its word. While this sits at Open, MR-08 fails the pull request.

Running the bridge twice over the same pull request left one record and one assessment, which is the whole point of the idempotency key. What no screenshot can show you is whether any of this is true — which is why the rest of this site publishes bundles you can download and verify instead.

What changes in your week

The control frameworks are not the hard part. Producing evidence somebody outside the team can check, on the day they ask, is the hard part.

If you own the control

Today

An examiner or internal audit asks how a change was approved. You open a ticket, a pull request, a chat thread and a screenshot folder, and rebuild the story of a decision made months ago from whatever people still remember. You attest quarterly to controls you have no practical way to observe in between.

With osfi-gate

The answer is a file. It was written the moment the change shipped, it names the policy version that applied, and it holds the scanner output the decision came from. You evidence every change instead of attesting once a quarter. A control the gate could not evaluate is reported as unknown, never as passing, so a gap surfaces in your own console before it surfaces in somebody else's finding.

If you run the engineering org

Today

Engineers get pulled off delivery twice a year to reconstruct approvals for work they have already forgotten. Someone maintains a spreadsheet of which repositories have which scanners. Every new compliance requirement arrives as a manual step a human has to remember, and eventually does not.

With osfi-gate

One job in the pipeline, beside the scanners you already run. No form to fill in afterwards and no dashboard to keep up to date. When risk wants to change what blocks a build, they change it in their own repository and you review the pull request. The argument moves from your build queue to a diff.

If you are asked whether we are ready

Today

Audit readiness is a project. It starts weeks out, pulls people off their work, and produces a set of assertions, each trusted because of who signed it.

With osfi-gate

The evidence already exists, because it was produced as the work happened. The person asking can check it themselves, offline, without going through your team and without trusting the tool that produced it. That is a whole category of "can you pull that for me" that stops arriving.

Change control, vulnerabilities, third parties, AI models

Change control

Evidenced on every change rather than attested once a quarter: independent approval, segregation of duties, protected branches, signed commits.

Vulnerabilities

Each finding is held against the SLA your policy sets for its severity, with the scanner output that produced it kept beside the decision.

Secrets and third parties

Secrets in code or in history, unpinned dependencies, unapproved registries, and licences you have ruled out.

AI models

The places your code calls a model, found in the code and not in a spreadsheet, joined to your inventory and checked against the E-23 tier you assigned it.

All 24 controls, and what blocks versus what only warns →

See a year of evidence, on a fictional bank

28repositories
1,418gate runs
Invented dataNorthwind Trust is not a real institution

A model that drifted from the version that was validated

The AI model registry: five detected call sites listed with their repository, file and line, provider, model identifier, region and model-risk inventory ID, above two alerts — one about a model that drifted from its validated version, one about a validation record with no matching call site.
Call sites are found in the code and joined to the model inventory, so the registry reports what is actually running. Two problems here that a spreadsheet cannot show you: a model now running a different version from the one that was validated, and a validation record being maintained for code the gate can no longer find.

Control status across the estate

The console's control status page: repositories grouped by team, each row a repository with a tick or cross under each control identifier, and the commit and age of the evidence on the right.
Every cell is a recorded decision, not a self-assessment. Two repositories are failing and one is stale, meaning its status describes an older commit — which the page says rather than quietly showing green.

Open the demo →

It opens with a short guided tour, and everything in it is invented. Prefer to look before clicking? Seven views, with captions →

What it will not do

These are on the front page, not in a footnote. Better you read them here than find out during an exam.

  • It does not assess your controls. The gate decides against a policy your second line owns and versions in Git. What this produces is the evidence of that decision.
  • It is not a vulnerability manager. It records that a finding existed, what the policy did about it, and when. Triage, ownership and remediation workflow belong somewhere built for them — feed the SARIF to that, and keep this for the evidence.
  • The gate decides; the console only explains. The console is a reader over signed evidence, not a second place where something can be enforced or overridden. If the two ever disagree, the bundle is right and the console has a bug.
  • Verification proves nothing was altered. It does not prove the decision was right. If the policy was wrong, you get a perfectly verifiable record of a wrong decision. That is why the policy is versioned, reviewed, and identified by commit in every bundle.
  • Coverage is of the enrolled estate, not of the institution. A repository nobody enrolled produces no evidence, and its absence cannot be seen from inside the evidence. There is a separate check that reports which repositories it can see, so you can compare that list against your own.
  • Every status is historical. It describes one commit at one moment, the last time the gate ran. It is not a live claim about what is shipping now.
  • Risk approves in Git, not in the app. A button in a web page would need its own signed audit trail before it meant anything, so there is no button.
  • Once evidence is locked, it cannot be deleted. Under compliance-mode retention a bundle cannot be removed before its retention date, by anyone, including you. That is the point of it, and it is a door that does not reopen. Choose your retention period carefully.

Free tools, no sign-up

They run entirely in your browser. What you type stays on your machine, is never sent to a server, and is never stored or seen by us.

E-23 AI Model Risk Alignment Tool →

Tier a model across seven risk dimensions in about five minutes. You get the tier, the rule that produced it so you can argue with it, the governance controls that tier requires with the E-23 section behind each one, and a model-inventory.yaml block ready to commit. Print it or save it as a PDF.

It tiers one model you already know about. The gate finds the ones nobody told you about.

Where this is, plainly

You are going to want to know whether this is a product or somebody's side project before you put it anywhere near your pipeline. Here is the honest answer.

  • One person builds it. Blake. Not a team, not a company with a support rota. That is a real thing to weigh, and you should weigh it.
  • The gate itself is finished and runs on its own repository. All seven milestones are complete, across seventy-one issues. The evidence this project produces about itself is produced the same way it would be for you.
  • Two things this list used to say were missing are now done. The reference policy lives in its own repository, so a policy change is a reviewable pull request where your second line already works; and the evidence bundles are published, so you can download a real one and check it without installing anything of ours. What is still open is named on the roadmap.
  • The WORM storage is tested against a real object store, not against AWS itself. The S3 Object Lock tests — compliance-mode retention, and legal hold as a separate lock — run on every change against SeaweedFS, an open-source store that implements Object Lock, and CI fails rather than skips if that store is not up. Until September 2026 they had never run at all. What no test here has done is lock an object in AWS S3: the store is a stand-in, and your bucket's configuration is yours to check.
  • It has not been through a supervisory examination. Nobody can tell you how an OSFI examiner will receive this evidence, and anyone who does is guessing.

If that reads as early, it is early. The upside of early is that a pilot shapes what gets built next, and you would be talking to the person who writes it.

Ask an awkward question

For a walkthrough, a pilot on your own estate, or a straight answer about any of the limits above. Email is the fastest way to reach me.

osfi@melx.buzz Connect on LinkedIn

Built by Blake. The roadmap says what is finished and what is not.