# Privacy and Disclosure Report

## 1. Scope

This note documents the public-release privacy posture of `spain-reference-personas-2025-v0.1`. It is a release-level disclosure report for a synthetic benchmark substrate, not a formal certified re-identification study against protected source microdata.

## 2. What is not published

The release does not expose the following as stable public fields:

- exact date of birth
- street address or exact municipality-level residence
- email address
- phone number
- document number
- stable public full-name column
- real employer name
- real school name
- direct contact identifiers of any kind

## 3. Structural safeguards in the public package

### 3.1 Synthetic identity and linkage

- Persons are keyed by `synthetic_person_id`.
- Households are keyed by `household_id`.
- Identifiers are synthetic release artifacts only.

### 3.2 Geographic abstraction

- Public geography is represented at region and municipality-class level.
- Exact local addresses are not published.
- Household context is represented through categorical abstractions rather than fine-grained real-world household records.

### 3.3 Household abstraction

- Household composition is encoded through counts and grouped household types.
- Economic context is binned through tenure, burden, vehicle access, and consumption-constraint categories.
- Narrative views do not reveal exact household addresses or named co-residents.

### 3.4 Narrative safeguards

- Public Spanish views are derived from structured fields.
- Views are designed for prompt usefulness, not for real-person imitation.
- Release tests explicitly block internal enum leakage in the public narrative layer.
- Public views do not include contact details or precise personally identifying narratives.

## 4. Release audit summary

Measured from `v0.1`:

- Personas: `1,000,000`
- Low disclosure-risk rows: `829,478` (`82.948%`)
- Moderate disclosure-risk rows: `166,341` (`16.634%`)
- High disclosure-risk rows: `4,181` (`0.418%`)
- Low uncertainty rows: `662,826` (`66.283%`)
- Medium uncertainty rows: `200,489` (`20.049%`)
- High uncertainty rows: `136,685` (`13.668%`)

Interpretation:

- The released package is overwhelmingly tagged low or moderate on internal disclosure heuristics.
- A small high-risk tail remains visible in metadata so downstream users can review or exclude it in stricter workflows.

## 5. Residual risk and caveats

This release should still be treated cautiously.

- Synthetic data can preserve sensitive structural patterns even when it excludes direct identifiers.
- Narrative views may feel realistic enough to encourage over-interpretation of individual records.
- Disclosure-risk tags are heuristic metadata, not a legal certification of zero re-identification risk.

## 6. Recommended handling for downstream users

- Use aggregate or subgroup analysis rather than point claims about any single synthetic person.
- Exclude `high` disclosure-risk rows when running especially conservative public demos.
- Use `population_weight` for macro summaries and publish methodology notes with any released findings.
- Avoid marketing, persuasion, or targeting claims framed as if the dataset were real survey microdata.

## 7. Relationship to evaluation and governance

This report should be read together with:

- [DATASHEET.md](DATASHEET.md)
- [SAFE_USE_POLICY.md](SAFE_USE_POLICY.md)
- [MODEL_AND_DATA_GOVERNANCE.md](MODEL_AND_DATA_GOVERNANCE.md)
- [EVALUATION_REPORT.md](EVALUATION_REPORT.md)
- [EVALUATION_METRICS.json](EVALUATION_METRICS.json)

## 8. Bottom line

`v0.1` is suitable for research, simulation, and benchmark development as a synthetic reference package. It should not be treated as a substitute for regulated disclosure review in high-stakes production settings.