Home / Who it is for

Underwriting model audit trail: decision records your actuaries and auditors can both read

If you own the models that rate a risk, set a price or push a submission into an underwriter's referral queue, you need a trail that shows which model version produced which output, and whether a person changed it afterwards.

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

The pack is due and the trail lives in four systems

You own the models that rate a risk, set a price, and decide which submissions go straight through and which land in an underwriter's referral queue. When a review arrives, from internal audit, from the reviewing actuary, or from a supervisory authority looking at a system you have classified as high risk, the question is narrow and awkward.

The question that is actually asked

Which model, at which version, produced this output for this submission, on this date, from which inputs, and did a person change the answer afterwards?

Most teams can answer it by hand, given enough people and enough notice, because the pieces do exist. The model version sits in a registry. The scored output sits in application logs with a retention window someone chose for operational reasons rather than evidential ones. The referral and the override sit in the policy administration system, described in an underwriter's free-text note. The rationale sits in a validation document that was accurate on the day it was written.

Four systems, and two readers who both need one of them

None of those four were designed to be read together by an actuary and an auditor at the same table. The actuary wants version, output and confidence per decision so the experience study holds. The auditor wants an ordered record whose integrity can be tested rather than asserted. When the two of them reconcile separate extracts, the meeting is spent on why the row counts differ instead of on what the model did.

That is the gap this page is about. An underwriting model audit trail is not more logging. It is one record, held per system, with an order you can recompute, so that reconstructing a single decision is a lookup rather than an archaeology project, and so that both readers are looking at the same rows.

If your question is broader than pricing, covering claims adjudication and fraud models across the book, the insurers overview in this section covers that ground. This page stays with the rating, pricing and referral models and the two people who have to agree on one record.

Pricing models move, which is the whole problem

A rating factor is refreshed. A challenger model is promoted after a quarter of shadow scoring. A segment is re-tiered mid-year. Every one of those changes is defensible on its own, and every one of them breaks a trail that records only the output. The evidence has to pin each output to the model identifier and version that produced it, at the moment it produced it, or the review turns into a debate about which version was live in March.

What one decision record holds

FieldWhat Auditrek storesWhat the reviewer does with it
Input hashA hash of the feature set your pipeline sent to the model, not the features themselvesShows the input set for that decision was not substituted, without putting applicant data in the ledger
Model id and versionThe identifier and version string your pipeline reports at inference timeTies a quote, a loading or a decline to one specific model build
OutputThe value your model returned, as your pipeline sends itReconstructs the priced or referred outcome for that submission
ConfidenceThe confidence or score your model reports, where it reports oneSeparates marginal cases from clear ones when sampling a period
Human override flagWhether a person changed the model's answer for that eventTurns the human oversight story into a count per system and per period
Subject referenceA pseudonymised identifier for the submission or applicant, applied by default at ingestionGroups the decisions that belong to one case without holding the identity behind them
TimestampThe event time carried on the record, stored and printed in reports; see the boundaries below for what the chain hash coversPlaces the decision inside the period under review, with ordering evidence coming from the sequence number
Sequence number and previous hashThe record's position in its system's chain, and the previous hash the record's own SHA-256 hash was computed fromLets the chain be recomputed server-side and returned as a tamper status rather than a stored flag

The workflow, from rating engine to report

  1. Register the model as a system and record the DPA

    Each pricing, rating or referral model is an ai_system in Auditrek. No API key and no webhook secret is issued until a DPA confirmation with a signed_at date exists, and that check runs server-side, so a pipeline handling applicant-derived data cannot begin sending events before the paperwork does.

  2. Point the pipeline you already run at the ingestion API

    Capture is a REST call or a webhook that accepts the inference event payload: input hash, model id and version, output, confidence, timestamp, human override flag. There is no model migration and no SDK to embed in the rating engine, so your model keeps serving quotes where it serves them today. Documented payload mapping guides cover AWS SageMaker and Azure ML, and anything that can make an HTTPS request posts the same payload.

  3. Let the minimisation gate refuse what should never be logged

    Ingestion inspects each payload for recognisable raw sensitive-field patterns, including social security number formats, ICD-10 style diagnosis codes and unredacted credit bureau fields, and rejects the request with a 422 and a clear error body before anything is written. Subject identifiers are pseudonymised by default. On medically underwritten life and health lines, that gate is the difference between an audit trail and a second copy of the health data.

  4. Watch the chain build append-only, per system

    Each accepted event takes the next sequence number and the previous record's hash for that system, and its own hash is computed as SHA-256 over that previous hash and the canonical payload, written in one transaction. The canonical payload covers the sequence number, input hash, output, confidence, human override flag and pseudonymised subject reference. Records are appended, never edited in place, so the chain for a pricing model is a single ordered history rather than a mutable table.

  5. Verify the chain instead of trusting a flag

    The verification endpoint walks the chain for that record's system from sequence one, 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 recomputation fails. An auditor gets a recomputed answer about the content and order of the records, not a checkbox someone could have set.

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

    The report generator produces PDF and HTML output for a chosen system and period, with its fields mapped to EU AI Act Article 12 and 13 fields. One regulation and one format set are available at launch. Every report carries an embedded, visible disclaimer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.

  7. Keep the dashboard between reviews

    The web dashboard shows decisions logged, systems connected, when the last report was generated, and tamper-verification status, written for a compliance officer who does not touch the pipeline. Access is auth-gated with multi-user org accounts, and people sign in with email and password.

What actuaries and auditors each get from the same record

One artefact, two readerships

The actuary reads model id, version, output and confidence per decision. The auditor reads an ordered chain whose integrity is recomputed on request. Both are looking at the same records for the same system and period, so there is no reconciliation of two exports that disagree by a few hundred rows.

The referral queue becomes evidence

The human override flag is a first-class field on the record and part of the hashed payload, so the proportion of model outputs a person changed in a quarter is a figure you pull from your own logged decisions, per system, rather than a narrative assembled from underwriter notes.

Version attribution survives model change

Because the model identifier and version travel on every event, promoting a challenger model or refreshing a rating factor does not blur the history. Decisions before the change stay attributed to the build that made them.

The production pricing path stays as it is

Auditrek records events your pipeline sends after it has produced its answer. There is no rewrite of the rating engine, no platform-specific SDK to certify, and no new dependency between a quote and a third party.

Applicant data does not become a second liability

An input hash instead of the feature values, pseudonymised subject identifiers by default, and an ingestion gate that rejects recognisable raw sensitive fields keep the evidence store narrower than the pipeline that feeds it.

A report you can hand over without overstating it

The output is scoped to one system and one date range, describes what that system did and when, and carries its non-legal-advice disclaimer on the face of the document. It documents your record-keeping; it does not assert that your practices satisfy any obligation.

Where the boundaries are

We would rather you know the edges before you plan around them.

  • What the hash covers today. The chain hash is computed over the sequence number, input hash, output, confidence, human override flag and pseudonymised subject reference. The event timestamp is stored on the record and printed in reports, but it is not part of the hashed payload in this release, and no cryptographic timestamp token is issued. Verification therefore attests to the content and the order of the records, not to the clock time written against them. If your review turns on proving when an event happened rather than in what order, say so early and pair the chain with your own time source.
  • Send confidence as a plain decimal. A confidence expressed in exponent form, such as 1e-7, can be re-read in a different string form and recompute to a different hash, which shows a record nobody touched as broken. Fixed decimal values avoid it until that path is corrected.
  • Long periods read better as HTML. The PDF writer currently renders to a single page, so a report covering many records is better read, and better circulated, as the HTML output while pagination is outstanding.
  • Connecting a system is a decision you keep. There is no disconnect or delete action for a connected ai_system in the app yet, so choose the first model you point at the API deliberately.
  • One regulation, one format set at launch. Reports map to EU AI Act Article 12 and 13 fields, in PDF and HTML. There is no Solvency II, NAIC or internal-model template in the generator today.
  • It records, it does not judge. Auditrek holds no opinion on whether a rating factor is appropriate, whether a model is accurate, or whether a price is adequate. Reports are descriptive and structural by design, and carry the embedded disclaimer.
  • It records what you send. If your pipeline does not report a model version or a confidence value, the record cannot invent one. Payload mapping is the first piece of work, and the guides for AWS SageMaker and Azure ML exist for that reason.
  • Nothing here is legal advice. Whether a given model is in scope, and what your own documentation has to contain, is a question for your counsel and your regulator, not for us.
  • Commercials and infrastructure. Billing runs through Stripe. The service is hosted on Amazon Web Services.

Articles 12 and 13, in the terms your filings already use

Article 12 of the EU AI Act deals with record-keeping and logging over a system's lifetime. Article 13 deals with transparency and the provision of information to deployers. Those are the two article numbers most underwriting and pricing teams meet first, because they are the ones that ask for evidence about a system rather than a statement about intent.

Auditrek's report generator maps its fields to those two articles for a system and date range you choose. That mapping is a documentation aid: it structures what your pipeline logged into the shape those fields describe. It is not a determination that your system, your logging retention, or your instructions for use meet the standard a supervisory authority will apply, and every report says so in its own footer.

Questions

Does Auditrek sit inside the quote path?

No. Your model returns its answer where it runs today, and your pipeline posts the inference event to Auditrek's REST endpoint or fires a webhook afterwards. Capture is a generic HTTPS call with a documented payload, with mapping guides for AWS SageMaker and Azure ML, so there is no model migration and no in-line dependency for the rating engine.

Can we log underwriter overrides and referrals, not just automated outputs?

Yes. The human override flag is one of the fields on every decision record, alongside the output, the confidence and the model id and version, and it is part of the hashed payload, so overridden decisions sit in the same chain as straight-through ones for the same system.

How does an auditor check the trail has not been edited?

Through the server-side verification endpoint. It walks the chain for that record's system from sequence one, recomputes each hash from the stored previous hash and payload, and returns a tamper status of valid or broken, plus the first broken sequence number when recomputation fails. It does not read back a stored flag.

Is the event timestamp covered by the hash chain?

Not in this release. The timestamp is stored and shown in reports but sits outside the hashed payload, so verification proves content and order rather than clock time. The boundaries section above sets out what the hash does cover and what to pair the record with.

What happens if a payload contains medical or credit bureau data?

The ingestion API inspects payloads for recognisable raw sensitive-field patterns, including social security number formats, ICD-10 style diagnosis codes and unredacted credit bureau fields, and rejects the request with a 422 and a clear error before any database write. Subject identifiers are pseudonymised by default.

Does the report cover Solvency II or state filings?

Not at launch. One regulation and one format set are available: PDF and HTML output for a chosen system and date range, with fields mapped to EU AI Act Article 12 and 13. The PDF currently renders to a single page, so use the HTML output for periods with many records. Every report embeds a disclaimer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.

What do we have to do before a pipeline can send events?

Record the model as a system and record a DPA confirmation with a signed_at date. Until that exists, the connect endpoint refuses to issue an API key or webhook secret, and the check is enforced server-side rather than in the browser. Note also that a connected system cannot yet be disconnected or deleted in the app.

Who in the team can see the records?

Access is auth-gated multi-user organisation access, and people sign in with email and password. There is no custom single sign-on, and no password reset or email verification flow in this release. The dashboard is written for a compliance reader: decisions logged, systems connected, last report generated, and tamper-verification status.

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.