Experiment with flag weight timeframes

[Undo1]

I'd like to run a simulation of what autoflagging looks like with reason weight only looking back n weeks/months/whatever. Curious if accuracy gets better (more responsive to reason changes) or worse (too responsive to reason changes).

[j-f1]

If it gets worse, you could split the reason weight into two parts — overall and recent accuracy. By adjusting the ratio between them, you could tweak responsiveness.