Cue-Dominant Extraction Explains Length Effects in Corrupted Reasoning Traces

Large language models increasingly consume externally generated reasoning traces in agentic and retrieval-augmented systems, yet these traces can be corrupted. Prior work reported that longer traces improve accuracy under corruption, but the mechanism remains unclear. Here we report seven controlled experiments (N = 1,990 inferences) distinguishing sequential chain integration from information survival hypotheses. Disrupting step order does not reduce accuracy and can significantly improve it (+7.5%, P = 0.001), contradicting sequential integration. Cue status is the dominant factor explaining position effects: at matched corruption (c = 0.4), final-answer cue integrity accounts for approximately seven times more variance than step position (31.7 vs 4.5 percentage points), with a secondary but measurable position effect. We propose a two-stage extraction model and discuss design implications for robust trace-based AI systems, where cue integrity—not trace length—should be the primary design target.