Across 21 speed-dating events, mutual yes decisions exceeded a model with fixed yes counts made and received by each person

We tested mutual interest in 4,184 recorded pairs from 21 speed-dating events at Columbia University in 2002–2004. There were 690 mutual yes pairs, or 16.49%. A model that kept each person's yes count made expected 709.8 pairs. The observed count was within its central 95% reference range of 681–739. A model that also kept each person's yes count received expected 632.7 pairs. Its reference range was 614–651. Thus, the added constraints changed the sign of the observed-minus-mean difference. The main excess was 57.3 pairs, or 1.37 percentage points. Seventeen of 21 events had an excess. All checks that removed one event retained a positive pooled excess. A 14-event check without selected study conditions gave a 1.55-point excess. Exact small-table checks, longer runs, a separate switch sampler, and a fresh-source replay supported the calculation. This is a conditional count benchmark for these records. The result does not establish a causal response to interest or transfer to other populations. The package contains code, source hashes, aggregate results, reference draws, figures, and reproduction instructions. Raw source data remain external under source terms.

speed dating · mutual interest · reciprocity · fixed margins · conditional randomization · binary matrices · reproducibility