Model cards¶

Every shipped model carries a stats file (*_stats_pt.json) with its calibration and conformal artifacts, produced by scripts/model-training/calibrate_model.py on the model’s held-out data. Methodology: see Statistical principles.

Model	Classes	Top-1	Top-3	Observed 90% coverage	Mean set size
Census 2000 surname	4	0.833	0.993	0.898	2.7
Census 2010 surname	4	0.808	0.984	0.898	2.8
Census 2020 surname	4	0.807	0.988	0.901	2.7
Florida voter surname	5	0.588	0.947	0.900	3.9
Florida voter full name	5	0.677	0.948	0.900	3.6
North Carolina voter full name	12	0.425	0.896	0.903	4.7
Wikipedia/Wikidata surname	13	0.775	0.907	0.901	8.1
Wikipedia/Wikidata full name	13	0.863	0.954	0.901	8.0
Wikipedia/Wikidata origin	90	0.626	0.809	0.900	24.4

Florida and North Carolina were retrained for 2.0 after an audit found source overlap in the earlier evaluation splits. The new artifacts split source rows before balancing and learn vocabulary only from training rows. The old numbers are not used as evidence.

Each JSON artifact contains the complete calibration record, including ECE, Brier score, temperature, all supported coverage levels, and reliability bins. The table keeps only the quantities most useful for choosing an estimator.

Set sizes show how much ambiguity remains. A Census estimate needs about 2.7 of 4 classes for 90% coverage; the 90-country origin model needs about 24. Last names do not identify fine-grained classes.

Dictionary estimators¶

Estimator	Categories	Source	Coverage	License
lookup_census_surname / lookup_census_first_name	6 census groups	Census 2000/2010/2020 surname + 2020 first-name files	surnames ≥100 occurrences (~90% of people); first names ≥100 (~94%)	Public domain
estimate_census_full_name	6 (4 on surname-model fallback)	census tables + census surname-model fallback	see `evidence_basis` column per row	Public domain
estimate_voter_file_full_name	5 (white, black, hispanic, asian, other)	six-state voter-file name dictionary (338k surnames, 136k first names)	registered voters, AL/FL/GA/LA/NC/SC	CC0

Dictionary probabilities are exact conditional frequencies. No calibration step is needed; their uncertainty is sampling error (Wilson intervals for the census tables) plus the naive-Bayes approximation for combined names (see statistical principles).

Evaluation populations¶

Metrics are not directly comparable across data sources. Census evaluation is person-weighted. Florida evaluation represents registered voters. North Carolina evaluation restores pre-deduplication voter frequency. Wikipedia and Wikidata evaluation represents unique notable people. Each stats file records the exact evaluation procedure.