Hightouch: Using Jev + Claude Opus for identity resolution review, cutting costs by 4x

Hightouch testing found that combining the decision model Jev with Claude Opus into a hybrid review approach can approach Opus's accuracy on identity resolution (IDR) over-merge detection while cutting costs by more than 4x. Used alone, Jev and GPT-6 Luna achieved accuracy of 88.8%–97.3%, with the decision model costing $0.5–$0.10 per 1,000 identities and responding in about 0.15 seconds, but on harder benchmarks it was about 6 percentage points lower than Opus. Jev's confidence calibration is better, with most errors accompanied by low confidence, so it performs better than Luna in the hybrid approach.