Human oversight evidence in the EU AI Act decision record
If you are looking for human oversight evidence under the EU AI Act, the hard part is rarely the policy wording, it is producing per-decision proof that a person intervened and that the record was not edited afterwards.
Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.
The row in your AI inventory that will not close
You own the AI inventory, and the line that stays open is human oversight.
Somewhere in the organisation there is a high-risk system with a person sitting in front of it: an underwriter who can decline what the model approved, a recruiter who can pull a screened-out candidate back into the pile, a reviewer who reads a score and then does something else. Your policy file already describes that person in the language your own filings use, human-in-the-loop override. What is usually missing is the per-decision record of the moments they actually intervened.
The question you are carrying is the one most compliance officers phrase the same way: can we prove it. Not can we describe the control in a policy, and not can we show a dashboard as it looks this morning, but can we produce, for one named system over one named date range, the decisions where a human overrode the model, the decisions where they did not, and a means for someone else to check that the set was not quietly edited later.
Practitioner commentary on Articles 12 and 13 puts the gap plainly: organisations can describe what their AI systems do, and far fewer can reconstruct how specific decisions were made, what data influenced them, and how human oversight was applied. Survey work published this year reports that only 33 percent of organisations have tamper-evident audit trails in place, meaning most cannot show that records were not modified after creation. The same commentary calls this a move from policy-based governance to evidence-based governance, and it reaches your desk as a request with a date range attached.
Auditrek is built for that request. The override flag is part of the decision record itself, inside the hashed payload, so human oversight evidence is a field in a chain rather than a narrative you reassemble under time pressure.
Where the override signal usually lives, and why that hurts later
In most stacks the override already exists. It is simply not attached to the decision.
It lives in the case management tool, where the row is editable and the timestamp belongs to the ticket rather than to the inference. It lives in a feedback table the ML team overwrites at the next retrain. It lives in an export someone produced for last quarter's gap assessment, now a file on a shared drive with no stated relationship to the model version that ran. None of those are dishonest. They are the wrong shape for an evidence request, because each one asks you to vouch for the joins yourself.
So the decision in front of you is a choice between two costs.
- Reconstruct later. Cheap until the moment it is challenged, then expensive in exactly the way that matters: you are the one explaining, in a room, why two systems disagree about how many reviews happened in March.
- Emit at the moment of the decision. This costs an engineering change in the service that already knows the answer, plus an honest conversation about what data may leave that service at all. In return the override flag, the model id and version, the input hash, the output and the confidence travel together in one record whose integrity can be recomputed by whoever you hand it to.
Auditrek is built for the second path and is deliberately narrow about it. It records what your system did and when. It does not decide whether your oversight arrangement satisfies any obligation, and it offers no view on that.
What each evidence shape can answer
| The question put to you | Reconstructed spreadsheet | Hash-chained decision record in Auditrek |
|---|---|---|
| How many decisions did this system make between two dates, and how many carried a human override? | A count you assembled, from joins you have to explain | A count over append-only records held per ai_system, with the human override flag in each record |
| Which model id and version produced the output the reviewer overrode? | Often absent, or inferred from deployment notes | Captured in the same payload as the override flag, input hash, output and confidence |
| Has this evidence been edited since it was written? | No mechanism beyond access controls and assurances | A verification endpoint recomputes the chain server side and returns a tamper status rather than reading a stored flag |
| Can we hand this to someone outside the team in a fixed form? | A CSV plus a covering explanation | PDF or HTML report output for a chosen system and date range, mapped to Article 12 and 13 fields, with a non-legal-advice disclaimer on every page |
The workflow Auditrek is built to support
- Register the system and record the DPA
An ai_system is the unit everything hangs from: the chain, the reports, the billing. No API key and no webhook secret is issued until a dpa_confirmation with a signed_at value exists for that system, checked on the server rather than in the browser. The mechanics of that gate have their own page.
- Decide what your override flag means
Before any code changes, settle the definition you would defend out loud. Is an override a reviewer changing the outcome, a reviewer declining to act on it, or both. Write it down once, because the field is a single flag on each decision record and the record cannot carry your caveats for you.
- Map your inference event to the payload
Ingestion is a generic REST call or a webhook, with documented payload mapping guides for AWS SageMaker and Azure ML. The payload carries input hash, model id and version, output, confidence, timestamp and the human override flag. No model migration is involved: the model stays where it runs today and the emitting service posts the event.
- Send the override with the decision, not after it
The flag belongs in the same payload as the output it qualifies. On write, Auditrek computes a sequence number and a previous hash from the last record for that ai_system, computes a SHA-256 hash over the payload, and stores the record append-only in one transaction. The override flag sits inside the hashed payload, not in a note beside it.
- Let ingestion refuse what should never have been sent
A data-minimisation gate inspects each payload for recognisable raw sensitive field patterns and rejects the request with a clear error before any database write, and subject identifiers are pseudonymised by default. For a reviewer override on a clinical or credit pipeline this matters more than it sounds: the evidence you need is that a human intervened, not the applicant's file.
- Verify the chain before you cite it
The verification endpoint walks the chain for that record's ai_system from the first sequence, recomputes each hash from the stored previous hash and payload, and returns a valid or broken status with the first broken sequence number. Run it yourself, so the integrity statement you make is one you checked rather than one you were shown.
- Generate the report for the range you were asked about
Choose the system and the date range and render PDF or HTML output mapped to EU AI Act Article 12 and 13 fields, designed to support documentation of those record-keeping and transparency obligations. The report generator has its own page, including what the output contains and the disclaimer carried on every page.
Why this shape helps the person answering the question
The override is evidence, not an annotation
Because the flag is a field in the hashed payload, the count of human interventions for a system and period comes from the same records as the count of decisions. You stop maintaining two numbers that have to be reconciled by hand before a submission.
Integrity is recomputed, not asserted
When someone asks whether the log could have been changed, the answer is a server-side recomputation of the chain that returns a tamper status, with the first broken sequence number if the chain does not hold. That is a mechanism you can describe in one sentence to a non-technical reader.
Your engineers add an emit, not a migration
Ingestion is a REST call or a webhook against a documented payload schema, with mapping guides for AWS SageMaker and Azure ML. The production inference path is not rerouted through Auditrek, which keeps the conversation with the platform team short.
One screen you can open in front of a colleague
The dashboard shows decisions logged, systems connected, when the last report was generated and the tamper-verification status, written for a compliance reader rather than an engineer, and auth-gated so colleagues in your organisation can see the same thing.
Sensitive fields are refused at the door
Inference payloads from credit, hiring and clinical pipelines are not ordinary web data. The ingestion gate rejects payloads carrying recognisable raw sensitive fields and pseudonymises subject identifiers by default, so the trail holds the fact of the override rather than the subject's record.
A fixed artefact you can hand over
Reports are generated for a named system and date range in PDF and HTML, and past reports are kept with when and by whom they were generated, so the document you cited in March can be reread in March's form.
Where the record stops and your counsel starts
Auditrek is record-keeping infrastructure, and the line matters enough to state twice. The product captures and chains what your systems did and renders it in a structured form. It does not interpret the Act for you, it does not assess whether your oversight arrangement is sufficient, and it does not tell you what you are required to do.
Every generated report carries a fixed, visible disclaimer stating that the report does not constitute legal advice and has not been reviewed by qualified regulatory counsel. Reports are descriptive and structural by design: field names, systems, dates, records, chain status. If you need a judgment on adequacy, that judgment comes from your counsel or your assessor, and it will be better informed if the underlying records are append-only and verifiable.
On data handling, the facts are short. Auditrek runs on Amazon Web Services in a single region, with no multi-region replication and no choice of data residency. Decision records hold an input hash, output, confidence, the human override flag and a pseudonymised subject reference; raw model inputs never arrive. People sign in with email and password, Stripe takes card payments and works out VAT at checkout, and Amazon SES sends account email. Sentry is used for error reporting.
Questions
What does Auditrek actually store as human oversight evidence?
A human override flag on each decision record, held in the same hashed payload as the input hash, the model id and version, the output, the confidence and the timestamp, stored append-only per AI system. Auditrek does not record the reviewer's reasoning or the subject's file: the field it keeps is whether a human intervened on that decision, bound into the hash chain so it cannot be edited without the chain failing verification.
Does sending override data mean changing our model or pipeline?
No model migration is involved. Ingestion is a generic REST call or a webhook from the service that already knows the outcome, against a documented payload schema, with field-mapping guides for AWS SageMaker and Azure ML. The production inference path is not rerouted through Auditrek, which is why this is usually a small change in an existing handler rather than a platform project.
How would I show that the override records were not edited later?
By calling the verification endpoint, which walks the chain for that record's AI system from the first sequence, recomputes each hash from the stored previous hash and payload, and returns a tamper status of valid or broken, with the first broken sequence number when the chain does not hold. The status is a recomputation at the time you ask, not a flag stored earlier, and the dashboard shows the most recent result per connected system.
Our overrides sit on credit and clinical decisions. What stops sensitive data entering the trail?
Two things, in order. No API key or webhook secret is issued for an AI system until a data processing agreement confirmation with a signed date is recorded, checked server side. Then the ingestion gate inspects every payload for recognisable raw sensitive field patterns and rejects the request with an error before any database write, and pseudonymises subject identifiers by default. The DPA gate has its own page if you want the mechanics.
Can I present a generated report as proof that we meet the EU AI Act?
No, and the report says so. Output is mapped to EU AI Act Article 12 and 13 fields and is designed to support documentation of those obligations, with a disclaimer on every page stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel. Whether your records and your oversight arrangement satisfy a given obligation is a question for your counsel or your assessor.
How is Auditrek billed, and what does it cost?
Billing is per organisation, tied to the number of connected AI systems, taken through Stripe. Starter is €1,500 a month for up to 2 connected systems and 50,000 decision records a month. Growth is €3,500 a month for up to 5 connected systems and 500,000 decision records a month. Both carry a 30-day free trial with no card to start. Sign-up is not open yet, so the step today is the waitlist.
Who in my organisation can see the records?
Access is auth-gated and scoped to your organisation, with multiple members able to sign in to the same organisation using email and password. The dashboard they see reports decisions logged, systems connected, the last report generated and the tamper-verification status, written for a compliance reader rather than an engineer.
Where do I read the field list the Act's record-keeping provisions concern?
Start with our field-by-field guide to Article 12 record-keeping, which covers what a log is expected to contain and the questions that follow, and the report generator page for what the generated document holds. This page stays on one narrower point: carrying the human override flag inside the hash chain so the evidence of intervention is a record rather than a reconstruction.