F1 Metric Misses Costly Merge Errors in Corporate Entity Resolution Systems

A software developer building an entity resolution layer for corporate and trade registry data found that tuning models to maximize the F1 score led to dangerous false positives, where two distinct companies were incorrectly merged into a single record. Unlike false negatives, which create fixable duplicates, false positives produce structurally blended entities that appear valid to all downstream systems, making the error nearly invisible. In one real case, a South Korean steel manufacturer was incorrectly merged with a logistics subsidiary simply because they shared part of a name and a city, and the error only surfaced when a user noticed an implausible business combination. The proposed fix replaces a single match threshold with three bands — auto-merge, review queue, and auto-reject — calibrated on the actual cost of false positives rather than test-set performance. On a corpus of East Asian corporate entities, this banded approach reduced the false positive rate by 73 percent while increasing review queue volume by only 12 percent.