Small bets with explicit goals. The failed ones stay visible.
Reduce context loss between research sessions without creating documentation overhead.
Test whether explicit metadata can beat semantic similarity on a small personal task set.