Our summary
Business records can name the same company in different ways, while similar names can mean different things. YTAL tested Jev on saved records for 525 new entries in a network of linked information. It tried eight matching designs, comparing their decisions with the old system rather than answers checked by people.
How answers were grouped changed the apparent uncertainty. Using saved rating answers and the same cutoffs, 13 percent of candidate pairs were left undecided. YTAL's rule required all other candidates to be ruled out before a match could proceed. Grouping those answers by entry left 39.6 percent of the 525 entries undecided.
Cheaper model calls did not guarantee a cheaper workflow. YTAL estimated that returning whole batches with undecided entries to the old system would add 4.1 percent to costs. It did not adopt this design. The records were not randomly sampled, and agreement with the old system cannot establish accuracy without checked answers.
Key takeaways
- Name what you count: a pair of possible matches is not the same unit as a complete entry with many candidates.
- Plan where undecided work goes. Sending a whole batch back can erase the savings from cheaper individual judgments.
- Compare against answers checked by people before claiming accuracy. Different decisions may be improvements or mistakes; old software is not a guaranteed source of truth.