Can Jev AI improve grading and marking accuracy?

For the last two weeks, everyone has been talking about a new AI model from a company called TypeSafe AI, which spent two years building what it markets as a new type of AI model: a "System One model" called Jev. You send text like with any LLM, but Jev doesn't return text or images in response. Instead, it returns decisions and probabilities. It's also very fast (TypeSafe claims 40 to 200 times faster than frontier LLMs) and a lot cheaper as well. This makes it a perfect choice for anything that needs a yes/no response, or a selection out of multiple choices. Companies are already using it for things like tagging support tickets, detecting spam, and, of course, playing Doom. I, naturally, decided to try out how well it worked for grading tests. Some numbers first I tried running Jev, as well as some open-source alternatives, on our internal evaluation set of representative questions and student answers. These are answers that have been graded by our AI models, and also reviewed and…