A useful target description explains which operations a device can execute and under what conditions their results remain meaningful. Qubit count is one part of that description. Connectivity, available gates, measurement behavior, reset, parallel execution, control timing, and calibration affect the experiment that a program actually becomes.
The local Bell simulator deliberately removes many of these constraints. It creates two qubits, applies supported mathematical operations, and samples a declared noise model. Moving its abstract circuit to a device requires checking a target's current capabilities rather than assuming that the simulator's instruction set is a hardware contract.
Different physical platforms impose different costs
Superconducting circuits, trapped ions, neutral atoms, spins, and photonic systems encode and manipulate quantum information in different ways. They consequently place different demands on refrigeration, vacuum, lasers, microwave control, optical routing, or other infrastructure. These categories contain substantial architectural variation. A technology label does not determine one fixed connectivity graph, gate duration, or error rate.
Coherence measurements also need context. For an approximately Markovian two-level system, T1 describes energy relaxation and T2 describes coherence decay, with 1/T2 = 1/(2T1) + 1/Tφ in the usual model. That relationship is useful when its assumptions apply. It does not predict the failure probability of an arbitrary scheduled circuit from two numbers alone.
A workload may spend time waiting between operations, experience crosstalk during concurrent activity, or encounter leakage outside the intended computational subspace. These effects can depend on the schedule and operating context. A device characterization must therefore describe what was measured, with which sequences, and how that evidence relates to the workload.
The target is more than a gate list
Suppose a target supports a controlled interaction only between certain pairs. The compiler must place the logical qubits accordingly or introduce routing operations. If interaction direction matters, an otherwise valid pair may require a different decomposition. Measurement and reset can occupy shared resources, so operations on different qubits need not run independently.
The same caution applies to dynamic control. Availability of mid-circuit measurement does not imply support for every conditional instruction, arbitrary classical processing, or a particular feedback latency. A capability check should inspect the relevant program shape and control operation, not merely a marketing label.
Record the target description used for compilation alongside the submitted artifact. When the provider exposes calibration or target revisions, retain those identifiers too. A later record of what the target supports today cannot fully explain a circuit compiled against an earlier configuration.
Worked example: routing and duration
Consider a synthetic line of three physical qubits, P0—P1—P2, with nearest-neighbor controlled-X gates in the needed directions. A program needs one interaction between logical qubits initially placed at P0 and P2. One simple routing plan swaps the state at P2 into P1, performs the interaction, then swaps it back to restore the mapping.
Under the textbook decomposition of each SWAP into three controlled-X gates, that plan uses seven controlled-X gates: three, then one, then three. If each costs a constructed 250 nanoseconds and all seven are sequential, their gate durations total 1.75 microseconds. This excludes single-qubit operations, buffers, measurement, and control overhead.
Now change the compiler's allowed output contract. If the final logical-to-physical mapping may change, restoring the mapping might be unnecessary. The same high-level interaction can then require fewer operations. The lesson is not that routing always costs seven gates; it is that placement policy and output mapping affect both resource use and evidence interpretation.
Operate against a validity envelope
Define an execution envelope with the constraints that matter to the workload: supported operations, permitted layouts, maximum scheduled duration, relevant calibration evidence, and any concurrency restrictions. Reject or recompile a program when those conditions fail. Repeating an unsupported schedule with more shots does not make it valid.
The envelope should be grounded in measured or provider-declared capabilities. The small noise slider in the teaching lab cannot certify such an envelope. Its purpose is to isolate relationships between a known channel, measurement statistics, and a decision rule.
Exercise: compare two compilation reports
Compiler A produces depth 20 and compiler B produces depth 16. A uses shorter native gates; B introduces a slower entangling operation and an additional measurement round. Which is faster, and which is more accurate?
Answer. Neither conclusion follows from depth alone. Compare scheduled duration with the relevant operation durations, then evaluate accuracy using workload-relevant evidence under comparable conditions. Depth measures dependency structure in a circuit representation; it is not elapsed time. The additional measurement can also change the experiment rather than simply lengthen it.
Compilation and portability develops this distinction further. Before that, run the first hybrid experiment and identify which hardware details the simulator intentionally abstracts away.