WORKED EXAMPLE · ACCEPTANCE RECORD 23 SEPTEMBER 2026

What your client receives at AI handover.

A signed acceptance plan, a report linked to the agreed criteria, and a clear record of what was checked and what remains missing. This worked example shows the documents we assemble for one acceptance decision.

ILLUSTRATIVE · FALSIFY PLAYS BOTH THE SUPPLIER AND THE CLIENT ROLE · DEMO SIGNATURES, NOT REAL PERSONS · NOT A BLIND TEST (SEE THE REPORT)

€3,000 fixed · One system · Up to three criteria · First draft within 3 working days of receiving your documents · details

1. The example at a glance

System
Credit-scoring classifier that recommends approve or decline on consumer credit applications (illustrative), trained on public data (UCI Statlog German Credit)
Version tested
Logistic regression trained by run_acceptance.py at seed 42
Criteria
C1 impact ratio for sex ≥ 0.80 · C2 accuracy ≥ 0.70
Decision recorded
Accepted: both criteria passed (0.857919 and 0.74)
Gap found
No hash of the trained model artefact (section 3)
Documents
Acceptance plan AP-2026-001 · Acceptance report AR-2026-001 · Change and re-review conditions CR-2026-001 · Verification statement VS-2026-001

2. The four documents

Below is what each document says. The signed files, their signatures and seal records are in the repository.

Acceptance plan · AP-2026-001

Signed by the client’s decision owner (Head of Credit Risk, illustrative) and the supplier’s delivery lead (illustrative), then sealed. The signed plan was sealed before the run-start records were registered. Scope: one system version, one acceptance round. Any later change to a criterion, threshold, test data or configuration counts as a deviation and needs a new plan version, signed and sealed again; the earlier version stays in the record.

IDMetricPass ifTest dataCriteria record
C1Impact ratio for sex (female ÷ male selection rate) on the held-out split≥ 0.80UCI German Credit, stratified 70/30 split, seed 425ce3c9c3484d…
C2Accuracy on the held-out split≥ 0.70same464df2bee6b4…

Rationale, written by the client’s decision owner (illustrative). C1: the system recommends credit decisions for natural persons; approval rates for women should not be substantially lower than for men. 0.80 is the four-fifths rule of thumb from bias-audit practice, not a legal standard, used as a floor for this round. C2: below 0.70 accuracy the recommendations would add more review work than they save. Falsify records this rationale and who approved it; it does not assess whether the metric or threshold is appropriate. Full text.

Acceptance report · AR-2026-001

CriterionPass ifObservedResult
C1 impact ratio (sex)≥ 0.800.857919PASS
C2 accuracy≥ 0.700.74PASS
Not a blind test. C1 repeats the configuration of our August 2026 worked example, so its value was known to us before the plan was written. The example shows how the record is built, not how a system performs.

Deviations from the plan: none. Decision: accepted. Signed by the supplier’s test lead and the client’s decision owner (demo keys). A failing result would have been recorded exactly as it came out. The report is the parties’ document: Falsify sealed its hash and linked it to the plan and to the system version as the record identifies it, by the training script’s hash and the seed, without a hash of the trained model (section 3). Falsify does not certify its content. Full text.

Change and re-review conditions · CR-2026-001

SituationWhat was agreed
Retraining on refreshed data from the same pipelineNo new acceptance round, as long as C1 and C2 still pass on the sealed test split; the supplier sends the client the new results within 10 working days
New data source or input features, new model family or configuration, new business process or applicant populationA new acceptance round against the same sealed criteria, unless both sides sign a new plan version
Production impact ratio below 0.80 for two consecutive months, or more than 25% of recommendations overridden in a monthRe-review of the acceptance within 30 days

A commercial agreement recorded at acceptance. Falsify records these conditions and does not monitor them. The agreement does not by itself make any change a “pre-determined change” under Article 43(4) of the AI Act. Full text.

Verification statement · VS-2026-001

Issued by Falsify OÜ. It lists the records by hash, the checks we performed and did not perform (section 4), and the gap we found (section 3). Registry receipts carry the registry’s technical signature; the statement itself is signed separately, here with a demo key. It was prepared with AI assistance and reviewed by Falsify OÜ. Full text.

3. A gap found during verification

The tested model is not identified by its own hash. The record identifies the training script and the seed, but does not include a hash of the trained model artefact. That limits how precisely the record identifies the tested model: someone who retrains it gets the same model only if the software environment matches. The gap is left open here, as the statement records it. In a real Pack we point out gaps like this while the documents are drafted, and the statement records whatever is still open.

4. What we checked, and what we did not

Checks we performedChecks we did not perform
Criteria hashes, test-data bytes and verdicts from the results
Run-final records chain to run-start and criteria records
Two valid signatures on the plan, the report and the conditions
Seal records list the exact bytes of every signed document
Every registry hash recomputes from the files
Every registry receipt signature (Ed25519) verifies
Every RFC 3161 timestamp verifies
The timestamps are in order: criteria, signed plan, run-start, run-final, report seal
The report refers to the sealed plan and both criteria records by hash
Whether the criteria are appropriate
Whether the results are correct beyond recomputing the verdict
Whether the tests ran after the plan (runner times are the supplier’s statement)
Whether other plans or runs existed
The identity of the signers (demo keys here)
Transparency-log inclusion proofs
A legal compliance assessment (section 5 shows where the record goes in the file)

5. Where this record goes in an AI Act file

Had this system been put on the EU market, it would fall under Annex III, point 5(b) of the AI Act: systems “intended to be used to evaluate the creditworthiness of natural persons or establish their credit score”. The requirements and provider obligations for these systems apply from 2 December 2027. For them there is no outside assessor: the provider follows the internal-control procedure (Article 43(2), Annex VI), in which it “examines the information contained in the technical documentation” to assess compliance, and keeps that documentation for ten years for the national authorities (Article 18). The provider is whoever develops the system, or has it developed and puts it into service under its own name (Article 3(3)); in a delivery that is often the client. The acceptance record is built to go into that file.

DocumentWhere it goesWhat it suppliesWhat stays with the provider
Acceptance plan, signed and sealedArt. 9(8): testing “against prior defined metrics and probabilistic thresholds” · Art. 17(1)(d): test and validation procedures · ISO/IEC 42001 A.6.2.4The metrics, thresholds, rationale and approvers, fixed and timestamped before the test runs were recordedWhether the metrics are appropriate to the intended purpose (Annex IV, point 4)
Acceptance report, dated and signedAnnex IV, point 2(g): metrics for accuracy “as well as potentially discriminatory impacts”, and “test reports dated and signed by the responsible persons”One dated, signed report per acceptance round, linked by hash to the plan, the criteria and the runs. C1 here is a discriminatory-impact metricThe testing itself and the test logs
Change and re-review conditionsAnnex IV, point 6: “relevant changes made by the provider to the system through its lifecycle”The agreed rule for which changes need a new acceptance round, and what triggers a re-reviewThe changes themselves, and any pre-determined change plan (Art. 43(4), Annex IV, point 2(f))
Verification statement, links and timestampsArt. 11(1): technical documentation · Art. 18(1): kept for ten years · ISO/IEC 42001 A.6.2.7What was checked and what was missing, and a way to show years later, offline, that the documents have not changed since they were sealedDrawing up and keeping the rest of the documentation

What only the provider can do. The provider draws up the EU declaration of conformity (Article 47) after the conformity assessment (Article 43). This record is evidence for that assessment, not the assessment, and nothing on this page says the example system complies. Risk management as a whole, data governance, human oversight, runtime logs and post-market monitoring are outside it. The full map, with the article texts, is in how PRML maps to the AI Act and ISO/IEC 42001; every Pack includes a one-page version for the client’s system. This is our reading of Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, not legal advice.

6. The record, in the order it was made

Times are the registry’s receipt times (UTC). Each record also has an RFC 3161 timestamp from timestamp.sigstore.dev and an entry in the public Rekor transparency log. The signed plan was sealed before the run-start records were registered; the runner times inside those records are the supplier’s statement.

TimeStepRegistry recordRekor index
11:50:23Criterion C1 recorded (impact ratio for sex ≥ 0.80)5ce3c9c3484d…122162600
11:50:27Criterion C2 recorded (accuracy ≥ 0.70)464df2bee6b4…122162688
11:50:42Acceptance plan signed by both sides and sealedda55fc4ac5cc…122163063
11:51:18Test run declared against C1 (run-start record)14bc38f2bd35…122163511
11:51:21Test run declared against C2 (run-start record)231970705630…122163513
11:51:37C1 result recorded (run-final record)86c224c760c7…122163530
11:51:41C2 result recorded (run-final record)74415372626b…122163532
11:52:20Acceptance report and change conditions signed and sealed86f76e446db7…122164037
11:55:30Falsify verification statement signed and sealedfbec1be1aa3a…122166615

7. Verify it yourself

  1. Clone github.com/studio-11-co/falsify and open examples/acceptance-record/.
  2. Run ./VERIFY.sh. It needs Python 3 with PyYAML, OpenSSL 3 (the LibreSSL shipped with macOS cannot verify these timestamp tokens) and OpenSSH 8 or later. No network is used.
  3. Eight checks run: criteria and verdicts, linkage chains, document signatures, seal records, registry hashes, receipt signatures, RFC 3161 timestamps, and their order.

8. What this shows, and what it does not

Who does what in a real Pack. You write the content and make the decision, and the responsible persons sign with their own e-signature tool. We draft from your existing documents, link, seal and point out what is missing. Arranging a document is not vouching for its content. Acceptance test preparation is not part of the Pack. If your team needs capacity to prepare the tests or the delivery file, we scope and quote that separately, in writing; the acceptance decision stays yours.

€3,000 fixed · One system · Up to three criteria · First draft within 3 working days of receiving your documents · details