We address the problem of measuring the dependency of multibiometric systems' scores, using Kolmogorov-Smirnov and Mutual Information criteria, and studying the validity of performance evaluation on chimeric persons. On the NIST-BSSR1 database, we formalize a common assumption in the literature: for independent scores, multibiometric systems can be evaluated on "random chimeric" persons. We show that this is not valid for dependent scores and propose a novel protocol for building "cluster-based chimeric" persons maintaining the level of dependency between scores. Finally, we show that performance evaluation for dependent modalities on such persons is equivalent to that obtained on "real" persons.