Abstract: We investigate whether rotating controlled fault challenges improves behavioral screening of poisoned client updates under non-IID federated learning. ChallengeShift-FL compares candidate models with the current global model on sensor-window interventions, applies a normal-data stability allowance, and maintains bounded temporal trust for soft aggregation. A reproducible synthetic experiment executes seven methods across seven scenarios and three paired seeds: 147 runs and 3,528 federated rounds. Comparisons include FedAvg, coordinatewise median, fixed and rotating behavioral weighting, the complete design, and two ablations. Rotation does not consistently reduce target-fault suppression. Under limited-knowledge adaptive poisoning, instantaneous rotation changes target-to-normal error from 54.22% to 53.78%, a paired difference of −0.44 percentage points; the complete design performs worse than its no-rotation counterpart. Under scaled replacement, the complete design yields 99.72% target-to-normal error, compared with 40.11% for instantaneous rotating weighting and 21.22% for coordinatewise median. Its stability allowance reduces honest suspicious flags but also admits damaging fault-suppression updates. All candidates retain positive aggregation weights, so zero hard rejection is structural rather than evidence of fairness. The findings expose failure modes of normal-only tolerance and trust inertia. They do not establish an industrial security guarantee or a general advantage of rotation. Challenge coverage and sample refresh remain confounded, and validation on public physical fault datasets is required.