`vally` experiments style testing, for doing "without " vs "with " (ie: baseline testing) to ensure we're actually improving LLM adherence and efficiency.

[richardpark-msft]

vally has an experiments mode, where you can do some baseline testing (amongst other things). It's a bit advanced since we're just trying to get our basic evals boostrapped, but will be essential as we get into "is this AI related feature we're building even improving things".