[experiment] self-improving evals

[mikecann]

It would be an interesting experiment to see if a model could be used to improve the guidelines iteratively.

So something like:

1. Model runs evals with current guidelines
2. Model observes output to ensure all evals still pass
3. Model makes a change to the guidelines (perhaps to remove something or refine to use less tokens)
4. Model returns to step 1

It might be important to use a "reasoning" model like o1 or r1 for this "architect" step before executing the evals again on a lower-model.