Comparison Methodology

We recommend fixed prompts, declared success criteria, and parallel runs.

• Define the task and audience
• Use identical prompts and settings
• Score outputs with a simple rubric
• Record date and model versions
• Re-run after provider updates