I wonder if multiple attempts at the opossum would produce better results.

If we didn’t have the previous example I would interpret this as pretty solid evidence that labs were training on the Pelican “benchmark”.

I just can’t imagine a model dropping so significantly from one version to the next on such a silly task.