Home / Who it is for

Credit scoring: an append-only record behind every automated decision

A credit scoring model audit trail is a per-decision record of which model version scored which application, what output came back, when it happened and whether a person overrode it.

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

What a credit scoring model audit trail has to contain

An audit trail for an automated credit scoring model is not the model's log file. It is a per-decision record: which model id and version scored which application, what output came back, what confidence sat behind it, when it happened, and whether a human overrode it.

Auditrek holds those records as a chain. Each record's SHA-256 hash is computed over the previous record's hash plus a canonical payload, which is the sequence number, the input hash, the output, the confidence, the human override flag and the pseudonymised subject reference. The record's creation time is stored with the record but is not part of that hashed payload today, which changes how you should describe the time evidence to a reviewer, and we come back to it below. Capture, chaining and append-only storage are verified against a running deployment: records ingested over both REST and webhook chained correctly from the genesis hash, each record's previous-hash field matching the hash of the record before it.

The verification endpoint walks that chain from sequence 1 and recomputes every hash in it. Change one of the hashed fields after the record was written and the recomputation stops matching, and the endpoint returns a broken status naming the first sequence number where it diverged. That is the mechanism, stated as built: it covers the fields folded into the hash, it is recomputed on request rather than read from a stored flag, and it does not cover the recorded timestamp.

If you searched for this, you are probably holding a version of the same situation. A scorecard or a gradient-boosted model runs inside the lending decision. Somebody has asked the question your team words as "can we prove it to the regulator": an internal auditor, a model validation function, a supervisory request, or your own board risk committee ahead of a readiness assessment. The evidence you can actually produce is spread across a feature store, a warehouse table that engineering has write access to, a ticket where the model version was noted by hand, and a folder of screenshots. It is not ordered, it is not tamper-evident, and the person asking wants a file covering a date range, not a live query you run in front of them.

The vocabulary this sits inside

Mid-market lenders tend to arrive at this from two directions at once. On the European side the words are record-keeping and logging obligations under Article 12 of the AI Act, and transparency and provision of information to deployers under Article 13. On the US side the words are model risk management and model governance, with the same underlying ask: show which model made which decision, and show that the record can be re-verified rather than simply asserted.

Whether your scoring system falls inside a particular classification, and what that classification then requires of you, is a question for your regulatory counsel. Auditrek does not answer it. What Auditrek does is hold the decision-level evidence in a structure those conversations keep asking for, so the answer is not blocked on reconstructing history from warehouse snapshots.

The decision in front of you

There are three usual routes, and it is worth being plain about the trade-offs.

Harden the warehouse. You add an events table, write a policy that says it is append-only, and rely on access controls. This is the cheapest route and the one that fails the specific challenge you are exposed to, which is not "do you have the rows" but "can you show the rows still hash to what they hashed to when they were written." Append-only by policy is not append-only by construction.

Build the chain yourself. A sequence number, a previous hash, a SHA-256 over the canonical payload, a verification job, a report renderer, and a retention story. None of it is hard. All of it becomes something your team now owns, maintains, and has to explain during a review, at the same time as everything else on the model risk roadmap.

Put a ledger beside the pipeline. Auditrek takes the inference event after your model has already returned a score. It is not a proxy in front of your endpoint and there is no model migration: the ingestion layer is generic REST and webhook, with documented payload mapping guides for AWS SageMaker and Azure ML. Deep native platform SDKs are not in the current scope.

The honest limits, so they are not a surprise later

  • Auditrek holds what you send it and nothing else. The hash covers the canonical payload fields, which are the sequence number, input hash, output, confidence, human override flag and pseudonymised subject reference, rather than every column in the row the event came from.
  • The record's creation timestamp is stored but is not part of the hashed payload today, so a timestamp edited in storage after the write still verifies as valid. Treat the recorded time as ordering evidence from the chain, record 412 was written after record 411, and not as cryptographically protected on its own.
  • There is no third-party trusted timestamp token on a record either. The plan named RFC 3161 timestamps; the column is reserved and nothing writes it today, so the proof of time is the chain's own ordering rather than an external authority's signature.
  • A chain walk detects a modified hashed field or a gap inside the recorded range. It is not on its own a proof that nothing was removed from the end of the chain, so reconcile the decision count against your own system of record for the same window.
  • The report generator ships with one regulation mapping and one format set at launch, and every report carries a disclaimer stating it is not legal advice and has not been reviewed by qualified regulatory counsel.

This is an evidence layer, not an opinion on your programme.

How one scoring decision becomes a record

  1. Register the scoring system, then record the DPA

    Each model becomes its own ai_system, so a champion and a challenger stay separate chains. No API key and no webhook secret is issued until a dpa_confirmation with a signed_at timestamp exists. The check is server-side: an unsigned system returns 403 on connect and cannot be bypassed from the client.

  2. Map your scoring event to the payload

    Six fields do the work: the input hash, the model id and version, the output, the confidence, the timestamp, and the human override flag. Mapping guides cover SageMaker and Azure ML event shapes; anything else posts to the same documented REST endpoint or webhook.

  3. Post the event after the score returns

    Auditrek receives decisions out of band, so the ledger is not a dependency of your lending path. A slow or unavailable ledger delays a record, it does not delay an applicant's decision.

  4. The minimisation gate inspects the payload before any write

    Server-side pattern validation rejects payloads carrying recognisable raw sensitive fields, including unredacted credit bureau fields and SSN formats, with a 422 and a clear error body before anything reaches the database. Subject identifiers are pseudonymised by default.

  5. The record joins the chain

    Auditrek computes the sequence number and previous hash from the last record for that ai_system, hashes the canonical payload with SHA-256 over prev_hash plus payload, and stores the fields atomically in one transaction. The hashed payload is the sequence number, input hash, output, confidence, override flag and subject reference; the creation time is stored beside them, outside the hash. Records are append-only per system; there is no edit path.

  6. Verify the chain, then generate the file for the window

    The verification endpoint walks the chain from sequence 1 and recomputes every hash from the stored previous hash and the hashed payload fields, returning a tamper status derived from that walk rather than a stored flag, along with the first broken sequence number if one exists. The report generator then produces PDF and HTML output for a chosen system and date range, mapped to the Article 12 and 13 fields, with the non-legal-advice disclaimer on every page.

What the ledger holds for a single credit decision

Recorded fieldWhat it holdsWhat a reviewer can do with it
Input hashA hash of the application features your model scored, not the application itselfMatch a copy you still hold against the hash, without the ledger ever storing applicant data
Model id and versionThe exact artefact that produced this outputTie a decision to a specific model version rather than to whatever is deployed today
OutputThe score, band or class the model returnedReconcile the ledger against the decision recorded in the origination system
ConfidenceThe confidence value your pipeline reportedSee where decisions clustered near a threshold across a date range
TimestampWhen the record was created, stored with the record but not part of the hashed payload todayBound a report to a review window, while treating the chain's ordering, not the stored time, as the integrity evidence
Human override flagWhether a person changed the automated outcomeEvidence for the human oversight question, at the level of individual decisions rather than policy documents
Subject referenceA pseudonymised identifier, applied by default at ingestionFollow one applicant's decisions through the chain without holding their identity in the ledger
Sequence number and previous hashThis record's position and the hash it is linked toRecompute the chain server-side over the hashed fields and get a tamper status, plus the first broken sequence if any

Why this shape suits a credit risk function

Your scoring stack does not move

Ingestion is generic REST and webhook with documented mapping guides for SageMaker and Azure ML. There is no retraining, no redeployment, and no rewrite of the inference service to accommodate an audit layer.

Bureau data stays where it already lives

The minimisation gate is designed to refuse unredacted credit bureau fields and SSN patterns at ingestion, returning 422 before a write. The ledger holds a hash and a pseudonymised reference, which keeps the evidence layer outside your most sensitive data footprint.

A tamper answer you can recompute, not a green tick

The verification endpoint recomputes each record's hash from the stored previous hash and the payload fields the chain covers, walking from the first record each time it is called. When a reviewer asks how you know the record set is intact, the answer is a mechanism they can re-run, not a boolean somebody could have set.

Written once, with no edit path

Records are append-only per ai_system. The sequence number and previous hash are computed server-side from the last record in that chain, so a decision cannot be quietly rewritten or reordered, and a gap inside the recorded range shows up as a broken sequence on the next walk.

Override is a field, not an anecdote

Because the human override flag is part of every captured event and part of the hashed payload, the question of how often people intervened in a given quarter is a property of the record set rather than a manual reconstruction.

One chain per model, including challengers

Records are chained per ai_system, so a challenger model running in shadow keeps its own ordered history and its own reports. The Starter plan covers two connected systems, Growth up to five.

A file for a window, in two formats

Pick a system and a date range and get PDF and HTML output mapped to the EU AI Act article 12 and 13 fields, with the non-legal-advice disclaimer embedded. The dashboard tracks decisions logged, systems connected, last report generated and tamper-verification status, so the person who needs that file does not need a query tool.

Questions

Does Auditrek sit in the path of a live credit decision?

No. Events are posted to a REST endpoint or a webhook after your model has already returned its output, so the ledger is not a dependency of the lending decision. If ingestion is unavailable, a record is delayed rather than an applicant.

Can I send raw bureau data or applicant identifiers?

The ingestion API is built to reject them. Server-side pattern validation inspects each payload for recognisable raw sensitive fields, including unredacted credit bureau fields and SSN formats, and returns a 422 with a clear error body before any database write. Subject identifiers are pseudonymised by default.

How does verification work, and which fields does it cover?

Call the verification endpoint for a record. It walks that system's chain from sequence 1 to the record and recomputes each hash from the stored previous hash and the hashed payload fields, which are the sequence number, input hash, output, confidence, human override flag and pseudonymised subject reference. It returns a tamper status of valid or broken, naming the first broken sequence number if it finds one. The status is derived from the recomputation, not read from a stored flag, and it speaks to those fields rather than to every column in the row the event came from. The record's creation timestamp is not among them.

Is the decision timestamp protected, or can it be edited after the fact?

Not by the hash today, and you should hear that from us rather than from a tester. The canonical payload covers the sequence number, input hash, output, confidence, override flag and subject reference; the record's creation time sits outside it, so a created_at value altered in storage still returns a valid tamper status. Records also carry no third-party trusted timestamp token: the RFC 3161 column is reserved and nothing populates it. What you can rely on for time is the chain's ordering, that record 412 was written after 411, together with your own system of record for the window. Binding the timestamp into the hashed payload is work we want to do; it is not shipped, so do not present it to a reviewer as though it were.

Does a report from Auditrek mean my scoring model meets its obligations?

No, and we will not say otherwise. The report is descriptive: it records what a system did, for which window, in the article 12 and 13 field structure. Every page states that it is not legal advice and has not been reviewed by qualified regulatory counsel. Whether your documentation is sufficient is a judgement for your counsel and your supervisor.

What has to happen before a scoring system can be connected?

A data processing agreement confirmation with a signed_at timestamp has to be recorded against that ai_system. Until it is, the connect endpoint returns 403 and no API key or webhook secret is issued. That gate is enforced on the server and cannot be bypassed from client state.

Which platforms are supported today?

A generic REST and webhook ingestion layer, with documented payload mapping guides for AWS SageMaker and Azure ML. There are no bespoke platform SDKs in the current scope; deeper native integrations are a follow-on, not something we are claiming now.

What will it cost, and how do people sign in?

Starter is €1,500 a month, covering two connected AI systems and 50,000 decision records a month. Growth is €3,500 a month, covering five systems and 500,000 records a month. Both carry a 30-day free trial with no card to start, and billing per organisation is tied to the number of connected systems through Stripe. Sign-in is email and password, with access scoped to your organisation. Auditrek is hosted on Amazon Web Services in a single region, with no multi-region replication and no choice of data residency.

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.