coverage: 92.654% (+0.3%) from 92.394%
when pulling 7f03f46 on shape
into ba7fbbf on master.

API looks quite clean!

Can you add a prefix or f-string or something for the name? T-strings might even be useful here.

Old way also still works like, if you need more flexibility on parameters, right?

T-strings might be useful in the future, but they are Python 3.14. I need to work with 3.12 for HPC reasons.

yes old way will still work.

I'm a bit confused looking at the implementation. The description shows:

def my_func(decisions:np.array[float]):
    ....

model.levers = [RealParameter(decisions, shape=(3,)),]

Several questions:

1. What is decisions here? In RealParameter(decisions, shape=(3,)), is decisions supposed to be:
   • The string "decisions" (the parameter name)?
   • The numpy array type annotation from the function signature?
   • Something else?
2. Where is shape implemented? I don't see shape being added to any Parameter.__init__() method in the diff. Is this still TODO?
3. Where are the bounds? The old API has RealParameter(f"l_{i}", 0, 1) with bounds. The new API example doesn't show bounds, are they still required?

Its wip, so I was doing stuff locally. The latest commit should provide some additional clarity, but there is still more work to be done.

1: The new API contained a mistake. I had forgotten the bounds.
2. "decisions" is indeed the name of the keyword argument. As allways with the workbench, parameter names must match keyword arguments on a function (if your model is a python function).
3. Shape is now added to all parameter classes. I have updated the samplers to reflect this change. Next, we need to turn the samples back into their original shape. After that, there is still the storage in the callback and, of course, the optimization side of things.

Thanks for explaining.

Might be useful to mark the PR as "Draft", to reflect its state.