Evaluation of AI research is important.

However, I doubt you can even figure out how to use a human-level explanation of the contents of an image to improve lossless image compression rate, so I'd suggest a different route.

There are already many small competitions in AI research - someone creates a dataset/task with a method of evaluating accuracy (usually agreement with human annotation), and others try to do better.

I'd like to see the use of open/common training and test sets eventually lead to a requirement that systems also be published and reproducible by others, or else the research papers about them are rejected. At that point, you can try whatever compression rate method you like on everybody's systems. I'd work toward that, rather than selling CRM.

CRM as you describe it here, if used to decide funding for AI research competitively, would encourage a lot of work in general data compression and not very much AI, at least for a few years.