Testing plan version 0.1 · amended once

Estimate whether a customer is 18 or over at unattended points of sale for age-restricted products, so that the machine either completes or refuses the sale. The harm the plan is written around is a minor being served; accuracy is judged against this purpose and nothing else.

Provider

Demo Provider

System

Point-of-sale age estimator

Annex III

1(b)

Classification

PROVISIONAL, pending legal. Not 1(a): the system establishes no identity and compares against no reference database, it returns an attribute. 1(b) is the nearest fit but is itself doubtful, since 1(b) is limited to sensitive or protected attributes and age is not a GDPR Art. 9 special category; Art. 3 also carves out categorisation ancillary to another commercial service. Recorded as 1(b) so the choice is explicit in every attestation rather than implied.

Sandbox

demo-001

Plan version

0.1The plan's own label, amended once since it was first signed. An amendment is a new digest, signed again; the label changes only if the parties change it.

What the provider and the regulator agreed before any run (Art. 57(5)). Every field that constrains a run is on this sheet, and the plan digest in every attestation commits to all of it: change a threshold and the digest changes.

Provider

Signed

Regulator

Signed

Opens

2020-01-01T00:00:00Z

Closes

2030-01-01T00:00:00Z

Max runs

10

Max run duration minutes

120

Max requests per minute

600

The gate refuses a run unless each committed digest is the pin.

Role

independent_heldout

Held by

test_lab

Groups

skin_tone_band, sex, age_band

Special category

Yes · Art. 4 bis

Commitment

sha256:5d43d03b9dad51c17c667cbb8cfe371e54f1d398dc2e0d43c786bcf2b3b6733a

Ledger

Committedledger seq 10 · 2026-09-28T07:22:34.912Z

A threshold here is the limit a run is judged against. It cannot be changed after a result is known: that is an amendment, and both parties sign it.

Type

accuracy_by_group

Article

15

Dataset

lab-heldout-v1

Metrics

mae

Groups

skin_tone_band, sex, age_band

Threshold · max_per_group

4.0

Min items per group

15

Confidence level

0.95

Interval method

bootstrap

Multiple comparison correction

bonferroni

Baseline

median_age Always predict the dataset median age. A model that does not beat this is not estimating.

Type

rate_by_group

Article

10, 15

Dataset

lab-heldout-v1

Metrics

false_adult_rate

Groups

skin_tone_band, sex

Threshold · max_per_group

0.01

Threshold · max_gap_between_groups

0.005

Min items per group

15

Confidence level

0.95

Interval method

wilson

Multiple comparison correction

bonferroni

Baseline

constant_is_adult Wave everyone through. The rate this produces is the floor any system must beat.

Type

decision_logging

Article

12, 19

Threshold · min_log_coverage

1.0

Threshold · max_unmatched_request_ids

0

Required fields

ts, request_id, output, model_id, model_version, latency_ms

Retention months

6

Each person signs in with their own organisation's identity provider; the plan decides which role that makes them.

Regulator

regulator@authority.example

Dpa

dpa@dpa.example

Test lab

lab@testlab.example

Provider

dev@provider.example

Data processor

annotation@labservices.example via test-lab-sso

Notified body

assessor@notifiedbody.example

activities:
- testing
sectors:
- 'Retail: unattended sale of age-restricted products'
goals:
- id: minors_not_served
  description: Show that the system refuses minors at a rate the authority accepts,
    in every skin-tone band and sex, on data the provider has never seen.
  evidenced_by:
  - minors_accepted_by_group
  - accuracy_by_group
- id: decisions_are_logged
  description: Show that every decision is logged with the fields Art. 12 needs.
  evidenced_by:
  - decision_logging
- id: human_oversight_design
  description: Discuss with the authority how staff override a refusal at the point
    of sale. Assessed by the authority; no automated test measures it.
cooperation:
- with: the data protection authority
  kind: authority
  purpose: Joins the review of the held-out set's legal basis before the first run.
- with: Independent test lab
  kind: expert
  purpose: Curates and commits the held-out set; never sees model outputs.
- with: Retail trade association
  kind: ecosystem
  purpose: Comments on how an override at the till would work in practice.

- activity: testing
  constraint: The provider never sees the held-out items or their labels, so it cannot
    tune on them; a failure is reported by group and metric only.
- activity: training
  constraint: No training happens in the sandbox; the model arrives trained.

regulatory_flexibilities:
- reference: AI Act Art. 14
  description: The authority gives guidance on the override design instead of a finding;
    the goal human_oversight_design is not scored.
information_exchange: A fortnightly call between the provider and the regulator; everything
  else through the console, whose actions are all in the ledger.

- id: signature_turnaround
  description: Working days from a plan amendment being proposed to both signatures.
  target: 10 working days or fewer
- id: exit_report_turnaround
  description: Days from exit to the exit report reaching the provider.
  target: Within the two months of draft implementing act Art. 6(4)

- reference: AI Act Art. 10
  note: Data governance for the held-out and stress sets
- reference: AI Act Art. 12
  note: Automatic logging of decisions
- reference: AI Act Art. 14
  note: Human oversight at the point of sale; guidance only
- reference: AI Act Art. 15
  note: Accuracy and robustness
- reference: AI Act Art. 4 bis
  note: Special-category data for bias detection
- reference: GDPR Art. 9
  note: Face images and group labels

- reference: AI Act Art. 4 bis
  challenge: Whether bias detection justifies processing skin-tone labels when synthetic
    faces cannot show the disparity.
- reference: AI Act Art. 14
  challenge: What human oversight means at an unattended point of sale.

processed: true
dpa_involvement: required
categories:
- face images (biometric data, GDPR Art. 9)
- skin-tone band (racial or ethnic origin, GDPR Art. 9)
- sex
- true age
legal_basis: AI Act Art. 4 bis(1) for the special-category group labels used to detect
  bias; PENDING LEGAL for the rest of the processing. Placeholder, like the thresholds.
note: The held-out set holds real face images labelled with skin-tone band and sex.
  The DPA named under roles is associated with the sandbox (AI Act Art. 57(10)).

safeguards:
- The model runs with no route to the data, the network or persistent storage.
- Labels and groups never leave segment A; only metrics and hashes leave.
- The regulator or the DPA can halt a run or suspend the participation at any time.
- Data keys are destroyed at exit.
serious_incident_procedure: Either party that suspects a serious incident (AI Act
  Art. 3(49)) tells the regulator the same day. The regulator suspends the participation
  from the console, which records the reason in the ledger, and informs the DPA when
  personal data is involved.

confidential:
- /participant/provider
- /datasets/0/processors
note: The provider's name until market launch; the test lab's subcontractor.

- id: demo-centre-undertaking
  kind: centre_undertaking
  centre: simulated-centre (histor, no Slurm, no Apptainer; model in a loopback-only
    netns)
  document_sha256: sha256:4f1e63cc76d2563609b8e56ba3bbc130489f7bd10b89ed7d31c01a90186b90f0
  valid_from: '2020-01-01T00:00:00Z'
  valid_until: '2030-01-01T00:00:00Z'
  clauses:
  - root_memory
  - crash_dumps
  - core_files
  - monitoring
  - hardware_service
  note: The template in docs/templates/, standing in for a signed undertaking of a
    simulated centre.

- id: prEN 18229-2
  title: 'AI Trustworthiness Framework — Part 2: Accuracy and Robustness'
  status: draft
  alignment: tracked

centre:
  name: simulated-centre (histor, no Slurm, no Apptainer; model in a loopback-only
    netns)
  partition: sandbox
  cpu_arch: x86_64
  runtime: apptainer
  personal_data_accepted: special_category