How it works

The framework parses a policy document into testable constraints, then probes the model and grades responses. Only metrics that pass an attribution test are kept.