Don’t trust us. Check.Try to break a sealed record
Coriqo
AI trust infrastructure for decisions about people

The assessment says a human reviews every significant decision. Show the auditor who, and for how long.

California’s ADMT rules and the UK’s Articles 22A–22D ask for the same thing: proof that a named person reviewed the decision and could have gone the other way. Coriqo records each review as it happens, and your auditor checks it themselves.

Coriqo holds only the review facts, never data about the people the decisions are about.

Maps to CCPA ADMT rulesUK DUAA Art. 22A–22DEU AI Act Art. 9–15NIST AI RMF 1.0
Where privacy programs stall

The assessment is filed. The auditor, and the people the decisions are about, care what happened after it was written.

Three questions come back on nearly every automated-decision assessment. The policy document answers none of them.

The assessment freeze

The risk assessment describes January

The assessment describes the system as written. Three model updates later, nothing records the day they diverged.

Self-reported numbers

"A human reviews every decision" is your own claim

The auditor asks how you measured it, across how many decisions, and whether anyone could have edited it.

The data paradox

Proving review means touching the data you are protecting

Showing review normally means pulling real people’s records into the audit. Coriqo holds only reviewer, time, override and decision type.

The CPPA’s ADMT rules took effect 1 January 2026, with significant-decision duties from 1 January 2027. The UK DUAA’s Articles 22A–22D have applied since 5 February 2026.

Illustrative example · a sample employer and a sample auditor, not a customer

The auditor asked who reviewed a rejected application, and for how long. The answer was on record.

01 · The question

One rejection, two questions

An auditor picks one application the résumé screener rejected in March. Who reviewed it, and for how long?

02 · The answer

Written down when it happened

The privacy officer opens that decision’s review record: reviewer, time spent, agreed or not. Recorded at the review, not for the audit.

03 · The pattern

The quarter, not one case

Across the quarter: under two seconds median, no disagreement. A finding against the privacy office’s threshold was already open.

04 · The check

Verified without the applicant's file

The auditor checks it through read-only access. No applicant’s data was pulled in to prove it.

That is the difference between a policy that says a human reviews every decision and a record of who did. This sample runs on the shipped EU AI Act pack; the dedicated ADMT and Article 22 packs are in the build.

One record, every framing

Every audit asks in its own words. Answer each one from the same record.

Pick the domain the auditor is asking about. Each EU AI Act Title III requirement is answered from records you already hold.

EU AI Act domain
Article 14 of the EU AI Act says oversight measures have to exist and be evidenced. That means human involvement you can show, not a policy sentence about it.
Human oversight1 check
E-06Blocks sign-off if absent

Human oversight measures must be documented, including the ability for an overseer to intervene in, interrupt, or override the system’s output.

EU AI Act · Art. 14(1)EU AI Act · Art. 14(4)
Answered by 4 sealed records · last change event #2,208

Requirement text is Coriqo’s built-in EU AI Act checklist (Regulation (EU) 2024/1689, Articles 9–15). Evidence counts are sample data. Dedicated ADMT and Article 22 checklists are in the build and will run on the same record.

Oversight Proof

"A human reviews every decision." Here is who, for how long, and how often they disagreed.

A sentence in the assessment is a promise. Coriqo records how long each review took and how often the reviewer went the other way. Either the record agrees with the assessment, or the day it stopped agreeing is on file.

Why it holds up

The referee is not the player

The measurement goes into an add-only record your team can’t rewrite, and the auditor checks it with a public key.

Why it grows in value

Better models widen the gap

As accuracy rises, review becomes a formality. You can’t infer scrutiny from model quality. You have to measure the person.

Where it's heading

A filing-ready package, not prose

The aim is an evidence package rebuilt per decision type and version. Coriqo never writes the assessment or classifies for you.

coriqo.io / vendor_hr_screening / oversight / resume-ranker@v2Sealed
Human review record
Employment identity · résumé screening · 2026-02-01 → 2026-07-31 · sample data
Median time on decision
1.6s
Across 5,200 reviewed decisions
Concordance with model
100%
No reviewer departure recorded
Decisions in scope
5,200
6 reviewers · 6-month window
Threshold you set · finding opened

Concordance stayed at 100% for 90 days across the significant-decision classes. Coriqo opened a tracked finding against the threshold your privacy office configured. It reports the number; the judgment stays with your privacy office and legal counsel.

Owner · privacy office  ·  due 2026-10-01
Recorded as event #2,411 in an append-only chain
Merkle root 7a3f…c901 · Ed25519, coriqo-key-2026-01
Anyone can verify it with the public key alone, without a Coriqo server
 What Coriqo does (a monitor)What Coriqo will not do (a detector)
Output"Resume screening: 1.6s median, 100% concordance, 5,200 decisions, pre-use notice filed against v2.""This processing is a significant decision."
The claimA fact about what happened to the record.A classification of the processing itself.
Who judgesYour privacy office, your counsel, the regulator.Not us. That is not our place.
If it's wrongIt reports what was recorded, so a wrong figure traces back to a record you can check.Every false ADMT call becomes your regulator's first impression of you.
Also for consumer-facing deployments

The opt-out is a system you honour, not a page you published once.

Pre-use notice, opt-out and access rights each leave a trail. Coriqo records each one as a sealed event tied to the system version it applies to.

NoticeFiled against the version it describes. A model change is a recorded event.
Opt-outWho declined, and when. Sealed and checkable.
AudienceYour auditors, the CPPA, the ICO and outside counsel.
BoundaryCoriqo never decides whether a use is a significant decision, or scores a reviewer.
Shared foundation

The proof underneath: one verification core, with privacy checks on top.

The record, the seal and the public-key check are the same code in every Coriqo deployment. Privacy is a set of checks on top, not a separate product with separate guarantees.

How verification works  ·  How it's built

Shared

Append-only event log. Every approval, change and review is sealed with SHA-256. Change a record after the fact and the check fails.

Shared

Merkle proofs. A short proof lets an auditor verify one event without holding the whole record.

Shared

Time-boxed external portal. Read-only access for a reviewer outside your company, with every view logged.

Privacy-specific

EU AI Act checks, run automatically. The Articles 9–15 requirement checks (eu_ai_act.yaml), seeded by NIST AI RMF 1.0. Every new privacy workspace runs them from day one.

Privacy-specific

Privacy decision classes. Employment, financial services, housing, education, healthcare, and profiling without a significant decision. Each is governed on its own track.

In the build

Dedicated ADMT / Art. 22 packs. Built from the CPPA’s ADMT regulations and DUAA Articles 22A–22D. In the build, not yet shipped.

Coriqo Shield · early access

Your ADMT record covers the AI you run. Shield covers the AI your staff use.

Shield replaces personal details before a message leaves each Mac, and keeps no copy of what anyone typed.

See Shield

We're looking for privacy teams to shape this vertical.

The shared record and EU AI Act checks ship today. The ADMT and Article 22 checks are in the build. Tell us which audit question keeps coming back.

Talk to us