The article describes a benchmark for extracting genealogical data from Portuguese parish records using AI, focusing on structured JSON output of people and relationships.
The benchmark consists of 84 images from 17 parish books, resulting in 434 records across baptisms, marriages, and deaths, spanning the 17th to 19th centuries.
Scoring uses ordered and unordered metrics, with field-level partial credit based on Levenshtein distance for names and temporal distance for dates, and a record-type mismatch zeroes all fields.
Experiments compare single-pass extraction (T0) against an agentic harness (T1) on Gemini 3.5 Flash Lite and Gemini 3.6 Flash, showing T1 improves scores but at higher cost.
Older (17th-century) records score lower due to faded ink, dense abbreviations, and physical damage, with the 17th century showing a clear penalty across all model-environment pairs.
Agentic T1 helps decompose complex images (e.g., splitting pages, adjusting contrast) but can introduce errors like misreading dates (e.g., 1606 vs 1666).
Limitations include small dataset size, lack of independent verification, no repeated sampling, and private dataset to avoid training data contamination.