Two practical principles for guarding against data leakage are developing a baseline intuition (knowing the normal performance range of features) and eliminating errors through architecture rather than relying on manual review.