Signature Uniqueness

Signature uniqueness is the specific registration rule that prevents a model registry's headline count from being inflated by entries that describe the same underlying model more than once. A model's signature is defined by its target variable, its filter criteria, its training sample size, and its validated performance metric together, and two registry entries sharing an identical combination of all four are, by this rule, one model rather than two, regardless of whether they were logged separately during development or carry different internal labels. Enforcing signature uniqueness matters because a registry that does not enforce it will tend to accumulate duplicate entries over time, as retraining cycles, experiment branches, and internal reorganizations each generate a new entry for what is, on inspection, the same underlying predictive model, and a headline model count built from an unenforced registry systematically overstates the size of an institution's genuinely distinct predictive fleet, sometimes without anyone inside the institution having noticed the inflation occurred. Enforcing signature uniqueness on a recurring schedule, not only at the moment a registry is first assembled, is what keeps the count honest as an institution's modeling operation continues to evolve, retrain, reorganize, and expand into new asset classes and jurisdictions over time.