METHODS.md

Methods: Spain Reference Personas Frontier

1. Purpose

Spain Reference Personas Frontier is a synthetic reference population and benchmark substrate for AI systems operating in Spain. The release is designed for controlled evaluation, simulation, prompt conditioning, subgroup analysis and benchmark development.

The dataset is not observed microdata. It does not represent real individuals and should not be used as a substitute for field surveys, administrative microdata or domain-specific validation.

2. Design Principle

The central design decision is that a persona should be a package rather than a long biography. The release therefore separates:

1. stable synthetic structure;
2. household and economic context;
3. compact LLM-facing views;
4. mutable actor state;
5. benchmark tasks and evaluation metadata.

3. Population Universe

• Universe: adult residents of Spain, age 18+.
• Scope: Spain-facing systems, not a generic Spanish-language population.
• Minors: represented only as household context in this release.
• Public geography: limited to region and municipality class.
• Public views: Spanish-first.
• Co-official and immigrant-language repertoires: modeled as metadata.

4. Artifact Inventory

Artifact	File	Grain	Rows	Role
persona_core	persona_core.parquet	person	1,000,000	Stable synthetic adult structure.
household_core	household_core.parquet	household	536,741	Household composition and economic context.
persona_views	persona_views.parquet	person-view	6,350,524	LLM-facing renderings with token budgets.
actor_state_init	actor_state_init.parquet	person	1,000,000	Mutable simulation-state scaffold.
benchmark_tasks	benchmark_tasks.parquet	task	1,800	Tasks, splits, scoring targets and replay seeds.
source_registry	source_registry.parquet	source	11	Release-level source inventory.
field_provenance	field_provenance.parquet	field-group	13	Per-field provenance mapping.

5. Source and Provenance Layer

The release includes 11 source-registry rows and 13 field-provenance rows. Provenance is documented at field-group level rather than as a fully reproducible generation script.

Field group	Provenance class	Source ids	Use
Core demographics and resident structure	official_statistics	ine-censo-2025	Calibration and evaluation.
Geography, municipality class and urban-rural profile	official_statistics	ine-censo-2025	Calibration and evaluation.
Household structure, minors, tenure and housing burden	official_statistics	ine-censo-2025	Calibration and evaluation.
Education, labor status, occupation class and socioeconomic tier	official_statistics	ine-censo-2025, cis-barometro-feb-2026	Calibration and evaluation.
Catalan and Aranese language identity and use domains	institutional_survey	idescat-eulp-2018	Calibration.
Basque competence and public-use domain modeling	institutional_survey	eustat-euskera-2024	Calibration.
Galician competence and domain use	institutional_survey	ige-galego-2023	Calibration.
Valencian public and professional domain use	institutional_survey	gva-valencia-2023	Calibration.
Digital access, commerce, AI usage and platform intensity	official_statistics	ine-tic-2025	Calibration and evaluation.
Reading, streaming, live culture and leisure participation	official_statistics	cultura-habitos-2024-2025	Calibration and evaluation.
Issue salience, trust and ideological orientation	institutional_survey	cis-barometro-feb-2026, interior-elecciones-2023	Evaluation.
Modeled latent values and behavioral style axes	modeled_latent	scope-latent-model	Calibration.
Spanish-first bounded narrative renderings	deterministic_rendered_narrative	frontier-rendering-policy	Rendering policy.

6. Generation and Calibration Overview

The public documentation describes a release-level construction pipeline: household synthesis, adult assignment within households, geographic and life-stage allocation, education/work/income assignment, migration and language-domain assignment, digital/media and cultural-profile assignment, civic/political/consumer/value-axis assignment, weight calibration and disclosure tagging, persona-view rendering, actor-state initialization and benchmark-task generation.

The current public release documents the resulting artifacts and evaluation metrics, but does not expose the full generation pipeline. This section should therefore be read as a release-level methodology summary, not a reproducible generation script.

7. Validation and Evaluation

Validation is summarized in EVALUATION_REPORT.md and EVALUATION_METRICS.json:

• package integrity: 7 Parquet artifacts and companion documentation are listed in the release manifest;
• row counts: 8,889,089 total package rows across the core artifacts;
• composition fidelity: region share MAE is 0.022 percentage points;
• age fidelity: age share MAE is 2.95 percentage points;
• weight stability: weights range from 0.9889 to 1.0551;
• token-budget compliance: all public persona views pass declared limits;
• benchmark coverage: 9 task families and 4 split regimes are populated;
• disclosure metadata: high disclosure-risk rows are 0.418%.

Age calibration remains the main calibration gap in v0.1.

8. Privacy and Disclosure Controls

The public release does not expose stable direct identifiers or exact personal contact fields. The privacy report states that the package does not publish exact birth dates, street addresses, email addresses, phone numbers, document identifiers, stable public full-name columns, real employer names or real school names.

Public geography is limited to region and municipality class. Disclosure-risk metadata is included so downstream users can filter stricter public demos or reviews.

9. Limitations

This dataset should not be interpreted as:

• a survey;
• observed administrative microdata;
• a census replacement;
• a predictor of real individuals;
• an election or public-opinion forecasting tool;
• an authoritative source for policy decisions without external validation.

10. Recommended Reproducibility Workflow

1. Pin release version.
2. Cite DOI.
3. Record artifact checksums.
4. Use held-out splits for benchmark experiments.
5. Report view type and prompt budget.
6. Filter disclosure-risk rows where needed.
7. Publish model/system cards for downstream use.