Coriqo
Trust & security · how Coriqo actually handles your data

Nothing here asks you to trust us. It asks you to verify.

Coriqo is built for institutions that run a vendor security questionnaire on every system that touches model risk data. This page is that questionnaire, answered plainly — deployment model, cryptographic integrity, AI data handling, regulatory mapping, and an honest account of where an early-stage vendor actually is today, not where a brochure would put us.

Maps to SR 26-2FDIC / FFIECSR 11-7 (superseded)NIST AI RMFEU AI Act21 CFR Part 11
Deployment & data handling

Your data stays in your environment. We never become a place your model risk data lives.

Coriqo deploys into your private cloud or on-premise — including air-gapped environments with no outbound network access. Connectors that read your existing systems are read-only by design: they observe your model registries, storage, and pipelines, but the source system keeps ownership of the data at all times.

Live

Private cloud / on-prem

Runs inside your infrastructure, not a shared multi-tenant SaaS you have to trust blindly. Air-gapped deployment is supported — no outbound network dependency required to operate.

Live

Read-only connectors

Connectors observe MLflow, Hugging Face, S3/Azure/GCS/NAS, CI/CD, and cloud accounts. They never write back and never take ownership — your source systems remain the system of truth for the underlying assets.

Live

Schema-per-tenant isolation

Each tenant gets its own Postgres schema, not a shared table filtered by a tenant_id column. Isolation is enforced once, centrally, at the request layer — not re-implemented per query across the codebase.

Live

CSV / Excel import

Bringing your existing inventory in is a real dry-run-then-commit workflow with a four-eyes approval step and dependency-aware rollback — not a one-way irreversible upload.

Labeled stub in code

Live system connectors

A continuous live pull from your S3/Snowflake/registry is an explicit v1 stub today — the scan endpoint says so in its own response. Today's import path is CSV/Excel, not a live feed.

Live

Role-based access control

Eight roles from read-only examiner access to org admin. A consultant role can prepare work but is deliberately excluded from governance approval — segregation of duties enforced by role, not convention.

Cryptographic integrity

Don't take our word for it. Recompute it yourself.

Every governance event — an approval, a validation sign-off, a monitoring breach, a finding — is appended to a SHA-256 hash chain. Alter any record after the fact and the chain breaks immediately, detectably, without needing to trust Coriqo's database. Periodically, events are sealed into a Merkle tree and the checkpoint is signed with Ed25519. An examiner can download an inclusion proof and verify a single event — with the public key alone, offline, with no Coriqo server and no network connection required.

$ python verify_proof.py --proof proof.json --pubkey coriqo-pub.pem
[PASS] Merkle inclusion proof — event proven in chain
[PASS] Ed25519 checkpoint signature — signed by coriqo-key-2026-01
RESULT: VERIFIED ✓ — no Coriqo server required
01

Hash-chained by design

Every transition computes a SHA-256 hash over the event and the previous event's hash. Timestamps are deliberately excluded from the hash formula so clock skew alone can never break the chain.

02

Merkle-sealed checkpoints

Events are sealed into a Merkle tree at checkpoint time. Inclusion proofs are O(log n) — an examiner can verify one event without needing the entire history.

03

Ed25519-signed

Each checkpoint is signed. Verification needs only the public key — never a live connection to Coriqo, never our infrastructure being up or trusted.

04

Optional public anchoring

Checkpoints can additionally be anchored to Sigstore's public Rekor transparency log — a record outside Coriqo's own control entirely, for institutions that want that extra layer.

Said plainly, not oversold: the hash chain makes tampering detectable, not impossible — a database administrator with direct access could still delete trailing rows outright, which breaks the chain but is only caught when someone actually runs the verification check. That's why verification is a button in the product, not a buried API call, and why Rekor anchoring exists as an option outside our own control. Rows written before a hash-chain fix are honestly marked as such rather than silently claimed as verifiable.

AI data handling

Zero-retention. Grounded. Never the actor of record.

Coriqo uses AI to draft — never to decide, and never to touch your permanent record without a named person approving it first. You can run Coriqo with every AI feature disabled and lose no other capability; that's a design constraint, not a marketing line.

By design

No training on your data

Your data is never used to train models. Every AI feature is a stateless call through one gateway per request — not a fine-tuning pipeline.

By design

Named-person approval required

No AI output enters the immutable governance record without a human approving it. The permanent record always shows the approving person as the actor — never "AI."

By design

Citation-grounded, ID-verified

AI answers are grounded strictly in your own stored records. Citations are validated against real IDs in your data; an answer citing something that doesn't exist is rejected or retried, not shown. Any answer's citations can be re-verified against the sealed chain on demand — including that a cited event's hash still matches the record it names.

By design

Metadata-only audit logging

AI feature usage is logged for audit purposes by metadata only — never the message content itself — into the same hash-chained record as everything else.

By design

Fails honest, never silent

No API key configured returns a clean "unavailable," not a broken response — and never a hallucinated answer standing in for a missing one.

Uneven today

Committee minutes parsing

Of the AI features shipped, this one is the most weakly grounded — free-text parsing into structured minutes, without the same citation-validation guard the others have. Worth asking about directly if this feature matters to your evaluation.

Regulatory alignment

Mapped to the frameworks your examiner or review committee already cites.

Coriqo doesn't parse regulatory text and doesn't claim to. What it does is structural: enforce the workflow discipline these frameworks describe — segregation of duties, independent review, an evidentiary trail — and tag findings against the frameworks they relate to as a curated, human-reviewed cross-reference, not an automated legal interpretation.

FrameworkWhat Coriqo structurally supports
SR 26-2Model risk governance: a fixed lifecycle (draft → review → validated → approved → production → retired) with a hard-enforced rule that the reviewer can never be the model's own owner — evidence of "effective challenge," not a policy someone has to remember. Whether an asset is a model under §II is recorded as a determination — the answers given, the outcome, and the rule that produced it — and the revalidation interval is held per model with a reason, because §V prescribes no frequency. §I is explicit that the guidance sets no enforceable standards and that non-compliance draws no supervisory criticism, so what Coriqo records is your framework and your reasoning, not a compliance verdict.
SR 26-2 footnote 3 · generative and agentic AIThe guidance excludes generative and agentic AI from its scope, and says the bank's own risk management and governance practices should determine the controls for what it doesn't cover. Coriqo records that scope call and then governs those systems against the AI rulepacks the workspace installs — NIST AI RMF and the EU AI Act below — rather than claiming SR 26-2 reaches them.
SR 11-7 · OCC 2011-12 (superseded)Replaced by SR 26-2 on 17 April 2026, along with SR 21-8 for BSA/AML systems. Historical gaps stay tagged under the framework that was in force when they were raised, so an older finding still reads against the standard it was actually judged by.
FDIC / FFIECExaminer correspondence workflow (Q&A threads, document exchange, findings) and one-button examiner packages give your examiner a self-contained, dated, independently verifiable record rather than a scramble through email.
EU AI ActZero-retention AI handling, human-approval gating on every AI-assisted output, and an auditable record of who approved what — the transparency and human-oversight posture the Act expects of high-risk AI use, applied to Coriqo's own AI features by design.
NIST AI RMF 1.0The 19 Govern / Map / Measure / Manage categories are seeded as control objectives, so evidence and findings attach to a named category rather than a free-text label. This is the framework substrate the US healthcare AI governance stack sits on.
Joint Commission RUAIHThe five certification domains ship as a deterministic requirement checklist for clinical AI documentation — governance, data management, risk and bias reduction, monitoring and validation, transparency and training. RUAIH certifies a health system's AI governance program, so these are the checks a vendor's documentation must satisfy for its customer to pass a domain.
ONC HTI-1 DSIThe predictive decision-support source attributes a developer has to publish per algorithm — purpose and intended use, risks and limitations, development details, demographic representativeness, bias management, validation — cross-referenced onto the RUAIH domains wherever both cover the same documentation gap.
21 CFR Part 11Append-only, hash-chained event history with independently verifiable signatures maps to Part 11's electronic-record integrity and non-repudiation expectations, for institutions where that framework is in scope.

Note: this mapping is a cross-reference to help your team scope evaluation, not legal or regulatory advice. Your own compliance program is what evidences compliance — Coriqo provides the data structures and workflow discipline that make that evidencing possible.

Where we are today

We'd rather you find out now than during a questionnaire review after signing.

Coriqo is an early-stage company working with design partners. That means the core governance, hash-chain, and AI-safety architecture is real and shipped — but we haven't yet accumulated the third-party attestations a longer-established vendor would have. Here's the honest state, plainly stated.

Governance state machine, hash chain, and Merkle checkpointslive in production

Role-separated review, self-approval blocking, hash-chained events, and independently verifiable Ed25519-signed checkpoints are shipped and running today — not roadmap items.

Off-host backups and a tested restore pathlive

Encrypted database backups run nightly to a separate off-host bucket, and a monthly automated drill restores the newest one into a throwaway database — checking that every tenant schema and the tamper-evident chain survived. A backup we haven't restored from is treated as if it doesn't exist.

Scope determination and per-model validation cadencelive

The §II scoping determination and the per-model revalidation interval are both stored, audited, and returned by the API, and the overdue list already reports whether an interval was set by the institution or inherited from the platform's old 365-day default. Both are recorded from the model's own page, and a model nobody has scoped reads as unassessed rather than as in scope.

Independent security certificationson the roadmap

We do not currently hold SOC 2 or any other third-party security certification. This is a roadmap item, not a completed one — we will not claim otherwise on this site or in a questionnaire response, and we'll tell you exactly where that process stands if you ask.

Live system connectorsv1 stub in code today

Continuous live pulls from your registries and storage are not built yet — the scan endpoint honestly labels itself a stub. Data import today is CSV/Excel, with a real dry-run and four-eyes approval workflow.

Full alerting / on-call integrationroadmap

Monitoring breach detection is real and server-computed. Alerting today is a best-effort email to the model owner — no webhook, SMS, Slack, or paging integration yet.

Customer basestated honestly

We are working with design partners at this stage. We won't publish a customer count or logos we can't stand behind, and we won't imply scale we don't have.

Infrastructure transparency

Deployment-agnostic by design — we don't lock you into infrastructure we don't control.

Coriqo's backend runs on Postgres with schema-per-tenant isolation, and deploys into your private cloud or on-premise environment rather than a fixed multi-tenant hosting footprint we operate. That's a deliberate choice for buyers who need to keep model risk data inside their own infrastructure boundary and existing vendor-risk approvals.

Database

PostgreSQL

Schema-per-tenant, not row-level filtering — isolation enforced once at the request layer.

Deployment

Your environment

Private cloud or on-premise, including air-gapped. We do not require you to trust a shared Coriqo-operated cloud.

Documents

WeasyPrint

Server-side PDF generation for examiner packages; falls back to structured JSON if the PDF toolchain isn't installed — never a broken response.

Transparency log (optional)

Sigstore Rekor

Checkpoints can be anchored to the public Rekor log outside Coriqo's own control, for institutions that want that extra independence.

Backups

Off-host, drilled monthly

Nightly encrypted backups to a separate bucket with server-side encryption at rest; a monthly automated restore drill proves they actually restore.

We keep this section deliberately conservative — if a specific sub-processor or hosting detail isn't listed here, ask us directly rather than assume; your deployment's actual infrastructure footprint depends on where you choose to run it.

Need our full security questionnaire response or an architecture diagram?

We'll send a completed vendor security questionnaire (SIG-lite or your own format), a detailed architecture diagram, and answer follow-up questions directly from someone who can speak to the code — not a sales layer reading from a script.