I think there is an opportunity to do the following:
a) Expose a subset of the tuning/searching variables for a given query to the user with a simple user-interface, as a separate lab project…
…Power users can play with these options and provide feedback in some manner, implicit (clicks-out of Google page) or explicit (Ah! these set of variables is best for me). In addition to feedback, they may optionally provide some meta-info about the query (say, tags). Google will benefit from feedback-directed-learning.
b) Idea (a) can be applied to local searches (like Google co-op), such that the choice of variables is user/query specific; or applied to global searches; or both; or a hybrid, where popular local optimisations are propagated as global.