Consensus Loss Proves AI Can Be Both Accurate and Transparent

This section examines the function of PEAR's two loss terms, its link to linearity, and whether it generates trivial or tainted explanations. We demonstrate that PEAR-trained models preserve meaningful explanations by adding garbage characteristics to the data; random features are rarely evaluated as significant, and PEAR can even further reduce mistaken attribution in certain situations. We also show that consensus training encourages models to become more linear, both quantitatively (via decreased linear-fit error across input subspaces) and qualitatively (through decision-surface visualizations). However, models regularized by weight decay become more linear without matching advances in agreement metrics, so linearity alone cannot account for better consensus.