# Model and Data Governance

## Artifact Owner

Spain Reference Personas Frontier is maintained by Miguel Guerrero Manchón / apol.

## Release Type

Synthetic population and benchmark substrate for AI systems in Spain.

## Versioning

Each release should include:

- release id;
- release date;
- population reference date;
- dataset version;
- artifact row counts;
- checksum manifest;
- evaluation report;
- privacy/disclosure report;
- citation metadata.

## Release Lifecycle

Recommended lifecycle:

1. internal validation;
2. documentation update;
3. disclosure-risk review;
4. checksum generation;
5. Hugging Face release update;
6. DOI generation or DOI metadata refresh;
7. Zenodo archival snapshot;
8. public announcement.

## Data Classification

The dataset contains synthetic records and public documentation. It does not intentionally contain direct personal identifiers.

Public release constraints:

- no exact addresses;
- no phone numbers or emails for personas;
- no identity documents;
- no exact dates of birth;
- no stable full names;
- no real employer or school identifiers;
- geography limited to region and municipality class.

## Evaluation Governance

Each release should maintain:

- composition-fidelity metrics;
- weight-stability metrics;
- token-budget compliance;
- benchmark-family coverage;
- split integrity;
- disclosure-risk metadata.

## Known Risk Categories

- overclaiming synthetic data as real data;
- misuse for political persuasion or microtargeting;
- misuse for individual-level decisions;
- disclosure-risk tail rows;
- representational gaps, especially known calibration weaknesses;
- downstream hallucination when used in generative systems.

## Mitigations

- clear documentation that the package is synthetic and not observed microdata;
- visible safe-use policy;
- DOI-pinned release snapshots;
- checksum manifests;
- explicit evaluation reports;
- disclosure-risk metadata for stricter filtering;
- future release notes that do not silently mutate prior versions.

## Change Policy

Changes that affect row counts, schema, weights, benchmark splits, release dates, DOI metadata, license or privacy posture should be versioned as a new release rather than as silent mutation of v0.1.