Using a large language model to score, grade, or judge output, such as text, code, or another model's answer, in place of a fixed rule or a human reviewer.

Used in evaluation dashboards and vendor reports to make a subjective, non-repeatable scoring pass look like a measurement instrument, usually by attaching decimal places to it.

• Measurement Costume: Presenting a model's opinion as a precise number so it can be charted next to real metrics.
• Unrepeatable Scoring: The same input can receive different scores on different runs, while the report still calls the result a benchmark.
• House-Cleaning Reality Check: Like asking a friend to rate how clean your kitchen is, getting a different number every time you ask, and calling it a "hygiene assessment framework."

Dancing with the Stars (Season 27): judges raise numbered paddles that look like precise measurement, while the scores rest entirely on personal taste and shift week to week for the same quality of performance.

Using a language model to score or grade output instead of a fixed rule or a human reviewer.

“Our LLM-based assessment layer delivers objective quality scoring at enterprise scale.”

“We asked a chatbot to rate each summary from one to ten, and it gave the same summary different scores on different runs.”