Don’t trust us. Check.Try to break a sealed record
Coriqo
AI trust infrastructure · for financial institutions

Show your examiner what your AI did, who approved it, and that nobody changed it.

Your board and examiners don’t read model cards. They ask who decided, who knew when it went wrong, and what was done about it. Coriqo re-checks your AI controls on a schedule, not once a year, and builds a sealed package your examiner checks on their own machine.

Talk to us Take the exam-readiness checklist Run a sample agent case

Don’t trust us. Check.

Three people behind one AI decision, at a sample bank An illustrative trail. Approved by J. Rivera in Model Risk, who signed off what the agent is allowed to do. Paged: the on-call risk officer at 2:16 AM, for an alert rated 10 on the bank's impact scale, acknowledged at 2:21 AM. Checked by the examiner, months later, on their own machine without trusting Coriqo. SAMPLE BANK · ILLUSTRATIVE APPROVED BY J. Rivera, Model Risk Signed off what it may do, v4 9 Mar PAGED On-call risk officer Impact 10 · acknowledged 2:21 AM 2:16 AM CHECKED BY The examiner On their own machine, without Coriqo Months later
Who decidesA named person. Coriqo never decides.
PostureRead-only. Never in your serving path.
Who can checkYour examiner, offline, with a public key.
Reads againstSR 26-2, NIST AI RMF, EU AI Act

Your board and examiner will ask what your AI did. Answer from the record.

TodayWith Coriqo
InventoryA spreadsheet someone remembered to updateA read-only scan of what is actually running
Decisions“Check the vendor’s logs”Every input, tool call and output, recorded in order
MandateNobody wrote down what the agent may doA versioned mandate; a step outside it opens a finding
Exam prepWeeks of scrambleOne sealed package from the record, signed by two people
Credibility“Trust us”Your examiner checks it with a public key
Record · hand over · keep checked

Controls checked continuously, and a package your examiner can verify.

01 · Record

Every decision, approval and fix is chained and signed

Each approval names the person and the exact version they signed. Every event links to the one before it, and signed checkpoints seal the log.

02 · Hand over

A package your examiner checks on their own machine

One sealed package: inventory, lineage, sign-offs and oversight. Two people sign it before it is sealed. Your examiner runs the open-source verifier against a public key, with no Coriqo account.

03 · Keep checked

Controls re-checked on a schedule, not once a year

Retests run on a schedule, and a finding closes when one passes, with a written reason. Evidence that has gone out of date is flagged. Every night, a scan looks for AI nobody registered in your AWS or Azure account.

SR 26-2 replaced SR 11-7 as the interagency model risk guidance in April 2026. Its footnote 3 leaves generative and agentic AI out of scope, so no one hands you a standard for your agents. You set your own, and the record shows whether you kept to it.

Three ways an AI program goes wrong.

One in five organizations IBM studied had a breach tied to AI nobody approved.1

01

The AI you didn’t know you had

A read-only scan of your AWS and Azure accounts flags what nobody governs. How →

02

The agent that stepped outside its mandate

A finding with an owner opens the moment it happens. Watch one →

03

The sign-off that shouldn’t have happened

An approval on an open critical finding gets flagged, with the evidence attached. How →

1. IBM Cost of a Data Breach Report 2025.

Published figures

What finding out late costs.

181 days

before the average breach is even identified (IBM 2025)

$1.88M

saved when a breach is caught inside 200 days (IBM 2025)

$14M

average cost of one conduct incident at a financial firm, two to three a year (RepRisk 2026)

$2.2B

US fines for systems-and-controls failures in 2024 (Compliance Corylated)

With a record kept as you go, a gap shows up as a finding on your own record first, not in an exam letter.

Sources
One incident · a sample bank, illustrative

2:14 AM. An AI agent tries to move money it was never allowed to move.

The bank and people are invented. Every step is something the product does today.

2:14 AM

The agent steps outside its mandate

Coriqo flags it against the mandate Model Risk signed in March, and opens a finding.

2:16 AM

The on-call risk officer is paged

Moving money rates a 10 on the bank’s impact scale, so a person is paged now.

2:21 AM

She acknowledges it

PagerDuty is updated too. Her name and the minute go on the record.

Next morning

The finding has an owner

And a due date, so it can’t drift into next quarter.

Day 9

Fixed, retested, closed

The red-team test reruns and passes. The finding closes with a written reason.

Months later

The examiner asks when a human knew

2:16, 2:21, day 9. They check it themselves.

Watch an agent get flagged
Shaped by practitioners

Model risk people told us what was missing. We built it.

The initial big value lies in streamlining processes / workflow and creating systems of record. Instead of relying on shared drive storage and creating Excel workbooks and email chains, your platform could help streamline this.
Model risk practitionerHas run MRM at small and large banks
Asked 14 OctBuilt 15 Oct

What the rest of that message became

  • Attestations by model owners
  • A Model Risk Index you can tune
  • Days past due on every overdue finding
  • Reports for ERM, the board risk committee and audit
  • Yearly regulatory attestations, prefilled from past validations
You want credible challenge or effective challenge… don’t supplant [the validator].

Built: contradiction flags a person accepts or dismisses, with a reason.

Former bank MRM oversight lead · design-partner call
If you go into the small community bank that has three branches, they don’t have a clue, and they’re doing it on an Excel spreadsheet.

Built: import that spreadsheet and be on the record in an afternoon.

Former bank MRM oversight lead · design-partner call
Verify

Most AI tools ask you to trust their record. Coriqo lets you check it.

Every flag, finding and sign-off goes into a record that can only be added to. Your examiner checks it offline with a public key. If anyone changes it afterwards, Coriqo included, the check fails.

How the record is sealedTry to break a seal

Illustrative sample data. verify_proof.py runs offline with no Coriqo account.
Where to start

Start from what is on your desk.

An exam requestWhat your examiner will ask for

The SR 26-2 request list, what each answer needs to show, and a 30-minute readiness review.

Examiner record → A consent order or findingProve the fix is working

What failed, what changed, who checked it, and whether it has held since.

Remediation proof → A self-checkThe exam-readiness checklist

Twenty-five questions an examiner can ask about the AI you run. No form and no email address.

The checklist →
Posture

Read-only

Never in your serving path or a runtime dependency.

Identity

SSO and expiring examiner access

Every access is logged.

Deployment

Hosted, or your cloud

Hosted in the US, or in your own AWS or Azure account.

Data

Never trains models

No AI output enters the record without a named person’s approval.

Independent

Verify without us

A public key is all an examiner needs.

Checked by third parties. Each checkpoint is logged in Sigstore’s public Rekor log and time-stamped by DigiCert. How the record is sealed