Feature request: Make the model evaluate the fish and then evaluate the mirrored version of the fish. Pick the highest value as the orientation. What could go wrong? Perhaps ask the user before mirroring?

We definitely need a text-to-stickfish model here