Optimal in theory but totally uncookable.

Linear programming has become flexible enough to allow way more constraints. Take a thousand variables for each food type, and then not only impose constraints for each nutrient, but also for cookability, for example not allowing flour to be used in a much larger quantity than milk, if one desires. LLMs could probably figure half of this out.

Then, you could add binary variables for each food type and add big M-constraints to ensure these correspond to whether the food is actually used in the diet or not, modifying the objective to favor either variety or simplicity. One could then add constraints on these variables to ensure foods are not used in too small quantities (too large ones are even simpler).

If I didn't enjoy food too much, I would do this now, solve the MILP and strictly follow this diet. The Wikipedia article does not name any modern applications or improvements upon his principle, did. nobody actually follow through this at all as LP exploded?