Home / Who it is for

Algorithmic hiring audit trail: one hash-chained record per candidate decision

If an automated screening model ranks, scores or filters your candidates, an algorithmic hiring audit trail means one durable record per decision: which model version ran, what it returned, when, and whether a recruiter overrode it.

Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.

The question your ATS export cannot answer

You own the AI inventory. Automated candidate screening often falls under the employment and worker management category the EU AI Act treats as high risk, and Article 12 is the record-keeping and logging line you keep returning to in your own filings.

What you can actually produce today is usually a CSV from the applicant tracking system: candidate name, a score, a stage. It does not carry the model id or version, so you cannot say which build produced which score after a redeploy. It does not carry a human override flag, so you cannot separate the decisions a person reviewed from the decisions the model effectively made alone. And it is a mutable table, so nothing in it distinguishes the row as it was written from the row as it is now.

The job in front of you is narrow and unglamorous: produce per-candidate evidence of what the system output, from which version, at what time, and what a human did with it, and be able to show that the evidence has not been rewritten since. The constraint is equally narrow. You are not willing to route candidate names, national identifiers or anything from a background check into another vendor's database to get that evidence, and the ML team is not willing to change the inference path for it.

What one hiring decision record holds

FieldWhat it recordsCovered by the chain hash
Input hashA hash of the candidate input the model scored, computed in your pipeline before you post, never the input itselfYes
OutputThe score, band or recommendation the inference response returnedYes
ConfidenceThe confidence value the model reported, where it reports oneYes
Human override flagWhether a person changed or overrode the model output, taken from your recruiter workflow or ATSYes
Subject referenceA pseudonymised stand-in for the candidate, pseudonymised at ingestion by default, never the applicant's nameYes
Sequence numberThe record's position in that ai_system's chain, computed by Auditrek on writeYes
Decision timestampThe time recorded against the decision on the recordYes, under hash version 2
Model id and versionWhich screening model, at which version, produced this output, from your model registry or deployment configStored on the record

The workflow Auditrek is built to support

  1. Record the DPA, then connect the hiring system

    An ai_system cannot be connected until a dpa_confirmation with a signed_at value is recorded. No API key and no webhook secret are issued before that, and the check runs server-side on the connect endpoint, returning 403 if the confirmation is missing. This is deliberate: hiring payloads are the kind of traffic that needs the paperwork settled before the first request, not after.

  2. Map your screening output to the decision payload

    Ingestion is a generic REST and webhook layer, with documented payload mapping guides for AWS SageMaker and Azure ML. There are no bespoke platform SDKs in the current scope, which is the point: anything in your stack that can post JSON after a scoring call, including an ATS webhook or a post-inference Lambda, can write a record. The production inference path stays as it is and no model moves.

  3. Post one record per candidate decision

    Each POST computes the sequence number and previous hash from the last record for that ai_system, computes the record hash as SHA-256 over the previous hash and the canonical payload, and writes the fields in a single transaction. Under hash version 2 the canonical payload is the sequence number, input hash, output, confidence, human override flag, pseudonymised subject reference and the decision's recorded timestamp. The response returns the created record with its hash and sequence number, so your pipeline holds a receipt for every decision it logged.

  4. Let the ingestion gate hold the line on candidate data

    Server-side pattern validation inspects each payload for recognisable raw sensitive fields, including SSN-style formats, ICD-10 style diagnosis codes and unredacted credit bureau fields, and rejects the request with a 422 and a clear error body before any database write. Subject identifiers are pseudonymised by default, so send a reference rather than an applicant's name or email address: the mapping back to a named candidate stays in your systems.

  5. Verify the chain rather than trusting a flag

    The verification endpoint walks the chain for that record's ai_system from sequence one, recomputes each hash from the stored previous hash and payload, and returns a tamper status of valid or broken along with the first broken sequence number. It is a recomputation at request time, not a stored boolean someone could have set, and what it evidences is the content, the order and the recorded timestamps of the hashed fields.

  6. Generate the report for one system and one date range

    The report generator produces PDF and HTML output for a chosen ai_system and date range, with fields mapped to EU AI Act Article 12 and Article 13 headings, designed to support documentation of those record-keeping and transparency obligations. One regulation and one format set at launch. Every report carries a fixed, visible disclaimer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.

Why this shape fits hiring specifically

Candidate identity stays on your side

Pseudonymisation of subject identifiers is the default at ingestion, and payloads carrying recognisable raw sensitive fields are refused outright with a 422. The ledger is designed to hold the shape of the decision, not the person. Raw model inputs never arrive, only the hash your pipeline computed.

Evidence per decision, not per quarter

A quarterly summary tells a supervisory authority what you believe happened. A per-candidate record carrying the model id and version that produced the output tells them what the system did, for one pseudonymised subject reference, in sequence, at a recorded time the chain hash covers.

The screening pipeline does not change

Records are posted after the decision over REST or webhook. Auditrek is not in the inference path, so the recruiter-facing latency budget and the ATS vendor contract are both untouched, and nothing about your scoring service's response to the applicant tracking system moves.

Human override is a hashed field

Whether a person reviewed and changed the model output sits inside the canonical payload the record hash covers, which is the distinction that matters most when you are describing oversight arrangements for a hiring system.

One dashboard, scoped to your organisation

Access is auth-gated and scoped per organisation, and the dashboard is written for a compliance reader rather than an engineer: decisions logged, systems connected, last report generated, and tamper-verification status. Sign-in is email and password.

Cost tracks systems and record volume, not headcount

Billing runs through Stripe per organisation against the number of connected AI systems, with VAT worked out at checkout. Starter is €1,500 a month for 2 connected systems and 50,000 decision records a month; Growth is €3,500 a month for 5 systems and 500,000 records. Both carry a 30-day free trial with no card to start, so a seasonal hiring peak is a volume question you can read in advance rather than a surprise on the invoice.

What Auditrek does not do here

Being precise about the boundary is more useful to you than a broad claim.

  • It does not assess your model. There is no bias scoring, no adverse impact calculation, no fairness metric. Auditrek records what the system output and what a human did with it.
  • It does not tell you what you are required to do. Reports are structural and descriptive, mapped to Article 12 and 13 field headings, and every one of them states in a footer that it is not legal advice and has not been reviewed by qualified regulatory counsel. What the output means for your obligations is a conversation for your counsel.
  • It is not a notarisation. Under hash version 2 the decision's recorded timestamp is bound into the hashed payload, so altering it breaks verification. The proof is still Auditrek recomputing its own chain, and there has been no independent cryptographic review of the hash chain implementation, so describe it as a tamper-evident record you keep rather than a third-party attestation of time.
  • It does not reconstruct history. The chain starts at the first record you post for that ai_system. Decisions made before you connect the system are not in it.
  • It does not supply your paperwork. The connect gate records that a data processing agreement was confirmed, with a signer name and a document reference. Auditrek does not yet offer a standard data processing agreement of its own for customers to sign.
  • It has no third-party attestation yet. There is no independent security audit or certification such as SOC 2 or ISO 27001. Sign-in is email and password, with no single sign-on.

Infrastructure runs on Amazon Web Services in a single region, with no multi-region replication and no choice of data residency. If your audit trail question is financial rather than employment, the same ledger shape is described on the credit scoring page and on the underwriting and pricing models page in this section, and the high-risk requirements checklist covers the wider obligation set this record feeds.

Questions

Does Auditrek check my hiring model for bias?

No. There is no fairness metric, adverse impact ratio or model assessment in the product. Auditrek captures decision records: input hash, model id and version, output, confidence, timestamp, human override flag and a pseudonymised subject reference, chained with SHA-256 and stored append-only per AI system.

Do candidate names or identifiers end up in the ledger?

Subject identifiers are pseudonymised by default at ingestion, and the ingestion API rejects payloads containing recognisable raw sensitive-field patterns with a 422 before any database write. Send a reference rather than a name or an email address; the mapping back to a named applicant stays in your own systems.

Does it sit in the inference path of my screening model?

No. Records are posted after the decision has been made, over a REST call or a webhook. There is no model migration and no change to how your scoring service responds to the applicant tracking system.

My ATS is a vendor SaaS I cannot modify. Can I still use this?

Ingestion is a generic REST and webhook layer with documented payload mapping guides for AWS SageMaker and Azure ML, and no bespoke platform SDKs in the current scope. Anything that can post the payload fields as JSON after a screening decision can write a record, whether that is your own service, a webhook consumer or a scheduled job.

How do I show that a record was not rewritten later?

The verification endpoint recomputes the chain for that record's AI system from sequence one, rehashing each entry from its stored previous hash and payload, and returns a tamper status of valid or broken with the first broken sequence number. Under hash version 2 the hash covers the sequence number, input hash, output, confidence, human override flag, subject reference and the decision's recorded timestamp, so editing the time on a record breaks verification along with editing the output. It remains our own recomputation of our own chain, not an independent attestation, and no independent cryptographic review has been done.

What is in the report, and can I rely on it with a supervisory authority?

The generator produces a PDF or HTML report for one AI system and one date range, with fields mapped to EU AI Act Article 12 and Article 13 headings, designed to support documentation of those record-keeping and transparency obligations. Every report carries a visible disclaimer stating it does not constitute legal advice and has not been reviewed by qualified regulatory counsel. Whether it satisfies a given examination is a judgement for your counsel and the authority, not for us.

Before I connect a hiring system, what is required?

A data processing agreement confirmation with a signed_at value must be recorded against the system, with a signer name and a document reference. Until it is, the connect endpoint issues no API key and no webhook secret and returns a 403, and the check runs server-side so it cannot be bypassed from the client.

What would one screening model cost, and can I try it today?

Starter is €1,500 a month for up to 2 connected AI systems and 50,000 decision records a month; Growth is €3,500 a month for up to 5 systems and 500,000 records. Both include a 30-day free trial with no card to start, and VAT is worked out at checkout. Auditrek is at the waitlist stage today, so nothing is open for self-serve sign-up yet, and we would rather say so plainly. Join the waitlist and we will come back to you with the payload mapping guide for the platform you name.

Sources

Join the waitlist

Leave your work email and we will send the payload schema, the verification endpoint documentation and a sample Article 12 and 13 report before you talk to anyone. Pilot cohort conversations start with your own pipeline, not a demo script.

One email when it launches, and one now to confirm your address. Unsubscribe any time. Privacy.