Compliance dossier — training data (AI Act, Article 10)

Automatically generated skeleton from the dataset analysis. The "to be completed by the expert" blocks require human review (provenance, compliance judgment, mitigation measures).

1. Identity & purpose

System / model: _____
Dossier version: _____
Analysis date: _____
Dataset: credit
Intended purpose: _____

_To be completed by the expert — intended purpose and what the data is meant to represent_

2. Provenance & legal basis

Origin of each source, licences/contracts, and — for personal data — initial purpose and GDPR legal basis.

⚠️ Potential personal data detected: email, full_name — a GDPR legal basis is required for these columns.

_To be completed by the expert — provenance per source (separate flow §16), licences, GDPR legal basis_

3. Composition _(filled automatically)_

Volume: 510 rows, 10 columns.
Variables:

Column	Type
`id`	BIGINT
`age`	BIGINT
`gender`	VARCHAR
`postal_code`	BIGINT
`income`	BIGINT
`employment_years`	DOUBLE
`loan_amount`	BIGINT
`email`	VARCHAR
`full_name`	VARCHAR
`defaulted`	BIGINT

Declared sensitive attributes: gender, postal_code, age.

Detected PII columns: email, full_name.

_To be completed by the expert — geographic / contextual / behavioural scope_

4. Preparation _(automatic observations)_

Duplicates: 10 duplicate row(s) → deduplication step to be traced.
Missing values: income (25) → handling strategy to be documented.

_To be completed by the expert — transformation log: collection, cleaning, labelling, enrichment, aggregation_

5. Quality _(filled automatically)_

Completeness: 25 missing value(s) (income).
Accuracy / outliers: income (3).
Uniqueness: 10 duplicate(s).

_To be completed by the expert — representativeness and relevance to the purpose_

6. Bias

Automatic analysis

Characteristic	Measure	Target	Verdict
Completeness	99.5 %	≥ 95.0 %	compliant
Uniqueness	98.0 %	≥ 99.0 %	to address
Validity	99.9 %	≥ 98.0 %	compliant

gender — « defaulted »: disparate impact (ratio) 2.14 · statistical parity (gap) 16.2 % ⚠️

Group	Count	« defaulted » rate
`F`	250	30.4 %
`M`	260	14.2 %

postal_code — « defaulted »: disparate impact (ratio) 1.83 · statistical parity (gap) 14.0 % ⚠️

Group	Count	« defaulted » rate
`93200`	100	31.0 %
`59000`	107	24.3 %
`13008`	103	19.4 %
`69003`	88	19.3 %
`75001`	112	17.0 %

Judgment & measures

_To be completed by the expert — impact on health/safety/fundamental rights and detection/prevention/mitigation measures_

7. Gaps & limitations

Detected points to examine as potential gaps:

[medium] duplicates
[low] missing — income
[low] outliers — income
[high] pii — email
[high] pii — full_name
[high] bias — gender
[medium] bias — postal_code

_To be completed by the expert — identified and addressed gaps; out-of-scope uses_

8. Governance & traceability

Analysis tool version: 0.4.1

Dataset fingerprint (schema + volume): d03427e41c5697a1

_To be completed by the expert — responsibilities, audit log, versioning, maintenance/updates_