Separate the sources of experimental differences¶
When results differ, first check the program, units, bit order and settings that actually took effect. Once both runs refer to the same quantity, separate changes in the physical model from numerical approximations and random estimates.
| Source | Comparison | What to vary |
|---|---|---|
| Physical noise | Ideal and noisy models of the same program | Channels, probabilities or rates |
| Time-integration approximation | Refine the calculation and compare with an independent method or analytic solution | Fixed-step time resolution or applicable adaptive tolerances |
| Finite trajectory estimates | Increase trajectories and repeat with predetermined independent seeds | Trajectory count and seeds |
| Finite measurement sampling | Increase shots and examine frequency variation across runs | Shots and sampling seeds |
More shots do not remove dephasing or refine time integration. Tighter tolerances do not add measurement samples. A fixed-step result may record tolerances without using them for adaptive step selection; see the simulation method guide.
Make a comparison answer a specific question¶
Change one setting first, then examine the probabilities, expectations or candidate distribution relevant to the experiment. Stabilization under numerical refinement is a consistency check at a stated precision. Two methods can still agree while sharing an incorrect assumption.
For small systems, prefer an independent analytic or matrix calculation. Keep a runnable small reference when increasing the system size, and record the scope of the error checks. Finite van der Waals interactions and hard blockade describe different models; switching between them is not simply a change in numerical error.
The introductory error lesson gives separate teaching comparisons. The noisy Hybrid project checks evolution against an independent Lindblad equation, then examines readout error and finite counts separately.