Exploration: counterfactual accuracy

[dzeber]

Once implemented (#197), the counterfactual accuracy evaluation needs to be tested out against a variety of datasets/models in order to surface any issues or potential improvements. This can be run as a data analysis project using the datasets in the repo and added to the examples dir. This can be considered done with at least 2-3 dataset/model pairs, but more is fine.

Based on these results, we would also like determine a protocol for making an automated pass/fail decision based on this evaluation, ie. a rule of thumb for what a "good" result looks like. Otherwise, if the evaluation is found not to be a good fit for such a protocol, this should be documented.

• Road-testing against multiple modelling problems
• Protocol for automated decision-making, if appropriate.