Unification Quality

The unifier links records in two passes: exact deterministic rules (email; phone + surname), then a probabilistic pass (Splink, Fellegi-Sunter) for links exact rules can't see. Because the synthetic generator emits ground truth, accuracy is measured, not asserted — pairwise precision/recall over matched pairs, on every run.

Precision (full)

84.85%

Recall (full)

96.44%

F1 (full)

90.28%

Review-band pairs

3,376

Baseline vs probabilistic, head to head

The deterministic pass alone is the baseline the Splink pass must beat — and whatever the numbers say is what ships here.

Loading...
No Results

Recall by injected mess type

Every source record carries ground-truth tags for the mess applied to it (nickname, typo, changed email, shared household email…). Recall per tag shows exactly which failure modes each pass handles.

No Results

Cluster shape

Loading...
Loading...

Pairs scored between the review and auto-merge thresholds go to a clerical review file rather than into clusters — merging on maybes is how golden records rot. Full methodology, thresholds, and the committed eval report: eval/results/ and ADR 0004 in the repo.