Calibration estimates or adjusts the relationship between intended operations and observed physical behavior. It can concern control amplitudes, frequencies, timing, readout discrimination, or other device-specific quantities. Operating a workload requires more than knowing that calibration occurred: the relevant calibration must support the assumptions used to compile, execute, and analyze that workload.
A calibration record is evidence about conditions and a procedure. It is not a permanent promise that the device will behave identically. Temperature changes, control drift, maintenance, shared-resource activity, and configuration changes can alter that behavior. The important question is whether the run remains inside its declared validity conditions.
Attach assumptions to an execution context
The teaching runner labels each batch with a context and rejects evidence that mixes batch contexts within one run. This demonstrates a useful boundary: counts taken before and after a known change should not be silently treated as exchangeable observations from one fixed distribution.
The string itself cannot prove stationarity. Two batches carrying the same label might still differ physically. On hardware, a context needs links to the relevant target, control settings, calibration evidence, timing, and other conditions. The label is an index into that record, not a measurement of its validity.
Similarly, a rule such as “use calibration younger than an hour” is an operational policy rather than a physical theorem. Appropriate validity conditions depend on the workload's sensitivity, observed drift, available monitoring, and the consequences of an incorrect result. A high-impact experiment may need a workload-specific check close to execution.
Separate preparation and measurement calibration
In the Bell model, p controls one preparation channel and q controls independent symmetric readout flips. Changing q changes the expected raw score without changing the prepared-state overlap. A calibration procedure aimed at one component should not be credited with improving the other unless there is supporting evidence.
A readout correction also depends on the accuracy of its calibration model. Estimating q from a finite sample introduces uncertainty. If readout errors are asymmetric or correlated, a single symmetric-flip parameter can be inadequate even when its fit looks plausible on a limited calibration set.
Preserve the raw counts and the correction model separately. This permits later reanalysis without pretending that corrected values were directly observed. It also makes it possible to distinguish a change in measurements from a change in interpretation.
Worked example: a changed readout context
Consider the declared Bell model with p = 0.02. Its prepared-state overlap remains 0.985. At q = 0.01, the expected raw score is approximately 0.955894. In a constructed comparison, q rises to 0.03. The parity attenuation changes from 0.9604 to 0.8836, and the raw expectation becomes:
(1 + 3 × 0.98 × 0.8836) / 4 = 0.899446.
This is close to, and slightly below, the lab's 0.9 acceptance threshold. A correction based on the old q would no longer describe the new measurement process correctly. Pooling old and new raw counts would estimate a mixture of contexts rather than either context's score.
The operating response is to mark the change, assess the new conditions, and create a separately identified execution when appropriate. A recovery run records its relationship to the earlier run while retaining its own budget, seed, measurements, and disposition. Its result should be read from its evidence, not assumed to match a previous recovery demonstration.
Check validity where changes can matter
Compilation can depend on target properties available before a job joins a queue. Execution may occur later. If the target or relevant calibration changes during that interval, the service needs a declared response: revalidate, recompile, reroute, or stop. Which response is possible depends on the provider interface and available evidence.
A successful validation package can establish that required fields and internal relationships are consistent. It cannot manufacture a missing device snapshot or prove that every unobserved source of drift was absent. Keep those evidence limits visible when reporting a result.
Exercise: decide whether to pool
A run has 2,000 shots per basis under readout context A. After recalibration, another 3,000 shots per basis arrive under context B. Both jobs complete, and their point estimates are similar. May the analyst call this one 5,000-shot checkpoint under A?
Answer. No. The completed jobs and similar point estimates do not establish identical measurement distributions. Preserve two contexts and analyze them separately unless a justified model explicitly supports their combination. If estimating a mixture is the intended objective, define its weighting and uncertainty rather than relabeling all observations as context A.
Noise and physical error budgets explains the model components behind this example. Timing, feedback, and dynamic circuits considers validity conditions that must hold inside an execution, on much shorter timescales.