trajectory-judge: measuring what LLM judges miss when an agent reaches the right answer the wrong way

A controlled study of what LLM judges of tool-using agents actually detect, under conditions where the ground truth is known by construction. A synthetic support-desk environment with a scripted oracle gives known-correct runs; a six-type injector breaks one step at a known index, keeps each fault's clean parent, and records whether the environment outcome survived and whether the final reply changed. Five judges (a rule checker, an outcome-only judge, a step-rubric judge with two models, and a self-consistency ensemble) are scored on detection, localisation, typing, calibration and cost over 400 trajectories. Comparing each fault with its clean parent shows that recall can credit a judge with detection it does not have. Every raw verdict is committed, so all tables and figures rebuild offline with no model calls.