Observable-Dependent Population Retention in GHZ States Under Dephasing-Dominated Noise

We report a state- and observable-dependent retention advantage in operational fidelity proxies under dephasing-dominated noise across multiple quantum platforms. We do not claim enhanced off-diagonal GHZ coherence; rather, we compare measured observables with different channel sensitivities under identical physical noise: GHZ population fidelity Fpop = P(|00…0⟩) + P(|11…1⟩) (diagonal in the computational basis) versus an X-basis product-state control implemented as H⊗N → delay → H⊗N → measure. On IonQ Forte (5–20 qubits), parity visibility follows V(N) = 1.041 exp(−0.030N) with χ²/dof = 0.86 across 5–15 qubits. At 20 qubits, V departs by −8.2σ, consistent with correlated-noise onset; model comparison (AIC) favors a piecewise description with regime separation. On IBM Heron (3-qubit GHZ experiments), the fidelity advantage ΔF reaches 69.7 percentage points at maximum idle delay, with the population survival ratio G = 12.5× from raw counts without error mitigation. Three platforms with average two-qubit gate fidelities in the 94–98% range do not show a detectable advantage within our dataset, consistent with a graded performance dependence on hardware fidelity. The effect is GHZ-specific within our data: W-states and 1D cluster states show reduced fidelities (−12% and −38% relative to product controls), with no indication of protection. All fidelities are from raw shot counts; no ZNE, PEC, or post-selection was applied. These results motivate regime-aware benchmarks that treat the measured observable as a first-class experimental variable.