Each post is a discovery. IRDME is the method — the finding is the story. All results link to their pre-registered hash.
A full account of the T_MODEL experiment series: we started looking for topology to predict attention head importance, found that activation magnitude (OAN) is a strong gradient-free predictor for binary sentiment classifiers, confirmed that the standard Taylor gradient criterion fails across all three models tested, and discovered that OAN's boundary condition -- where it works and where it doesn't -- may be more informative than the positive result itself.