Versioned rubric templates for common model-backed evaluation tasks.
Templates provide stable IDs and criteria while allowing each assertion to choose its provider and pass threshold. The template version is recorded in metric results so historical runs retain the criteria they used.