narrow test using emasculated ai models. Some limited application however they didn't evaluate usefulness. Redo the test with models that use tools, have more skillful prompts.