Commissioning is the transition from assembled hardware to operational computational instrument. A device is not commissioned when the fridge is cold or the trap is stable. It is commissioned when the platform can repeatedly accept workloads, produce traceable results, expose health signals, and recover from ordinary disturbances.
For a systems organization, commissioning is not a ceremony. It is a controlled release.
43.1 Commissioning as release management
A new quantum processor, upgraded control stack, new trap geometry, new cryogenic module, or new provider target should pass through staged release gates. Each gate narrows uncertainty.
View diagram source
flowchart LR
Design[Design intent] --> Install[Physical install]
Install --> Bringup[Subsystem bring-up]
Bringup --> Characterize[Characterization]
Characterize --> Calibrate[Calibration]
Calibrate --> Validate[Validation workloads]
Validate --> Limited[Limited availability]
Limited --> General[General availability]The mistake is to move directly from Calibrate to General availability. Calibration only proves that selected controls can be tuned under known conditions. Availability requires repeatability, monitoring, user contracts, and fallback behavior.
IBM’s processor documentation distinguishes processor families and revisions, and Quantinuum’s H2 operational documentation exposes how architecture affects scheduling and transport. Those details matter because commissioning criteria must match the modality rather than pretend every QPU has the same topology or control semantics. [R79] [R80]
43.2 Commissioning scope
Commissioning should cover the full stack:
| Layer | Commissioning question |
|---|---|
| facility | are power, cooling, vibration, magnetic, and access constraints stable? |
| device | are qubits or modes usable across the intended operating region? |
| control | do timing, waveform, transport, and feedback paths meet tolerances? |
| calibration | can calibrations converge and remain valid long enough to serve users? |
| compiler | does the target model match actual instructions and connectivity? |
| scheduler | can admitted jobs run without violating health or fairness rules? |
| data | are all artifacts captured with provenance and versioned schemas? |
| support | can incidents be triaged without heroic specialist intervention? |
View diagram source
flowchart TB
Commission[Commissioning program] --> Facility[Facility readiness]
Commission --> Device[Device readiness]
Commission --> Control[Control readiness]
Commission --> Calibration[Calibration readiness]
Commission --> Compiler[Compiler readiness]
Commission --> Scheduler[Scheduler readiness]
Commission --> Data[Data readiness]
Commission --> Support[Support readiness]A commissioning plan that tests only quantum benchmarks will miss failures in access control, cost accounting, queue behavior, data lineage, and user support.
43.3 Gate design
Each gate should have measurable entry criteria, exit criteria, and rollback criteria.
commissioning_gate:
gate_id: string
owner: string
entry_criteria:
- prerequisite completed
- telemetry available
- rollback path tested
tests:
- name: string
threshold: string
sample_size: string
exit_criteria:
- metric within bound
- artifact archived
- reviewer approval captured
rollback_criteria:
- metric breach
- data loss
- unsafe operating stateView diagram source
stateDiagram-v2
[*] --> Draft
Draft --> Ready: entry criteria met
Ready --> Running: execute tests
Running --> Passed: thresholds met
Running --> Failed: breach
Failed --> Remediate
Remediate --> Ready
Passed --> Released
Released --> Monitor
Monitor --> Rollback: regression detected
Rollback --> RemediateA gate without rollback criteria is not a gate. It is a wish.
43.4 Bring-up experiments
Bring-up experiments are not user workloads. They are diagnostic probes. Their purpose is to expose whether the device, control stack, and analysis pipeline are consistent with the target model.
Typical bring-up families include:
| Experiment | What it probes |
|---|---|
| spectroscopy | frequency or transition landscape |
| Rabi / Ramsey / echo | control amplitude, detuning, coherence |
| readout assignment | measurement discrimination and bias |
| randomized benchmarking | aggregate gate quality |
| crosstalk probes | interference between controls |
| topology probes | whether compiler connectivity assumptions hold |
| stress jobs | queue, thermal, drift, and artifact pipeline behavior |
Historical Qiskit Experiments releases documented pulse-calibration experiments and schedule and parameter management. That subsystem was removed following the removal of Qiskit Pulse in Qiskit 2.0; it should be distinguished from the project’s current characterization and verification features. Commissioning tests must use the target platform’s currently supported control interface. [R84] [R85]
View diagram source
flowchart LR
Probe[Diagnostic probe] --> Extract[Extract parameter]
Extract --> Store[Store calibration artifact]
Store --> Bind[Bind to target model]
Bind --> Validate[Validate with benchmark]
Validate --> Decision{Accept?}
Decision -- no --> Retune[Retune or repair]
Retune --> Probe
Decision -- yes --> Promote[Promote to release candidate]43.5 Commissioning data room
A serious commissioning program maintains a data room. This is not a slide deck. It is a structured evidence repository.
View diagram source
flowchart TB
DataRoom[Commissioning data room] --> Design[Design intent]
DataRoom --> Config[Configuration snapshots]
DataRoom --> Cal[Calibration history]
DataRoom --> Bench[Benchmark records]
DataRoom --> Incidents[Bring-up incidents]
DataRoom --> Waivers[Waivers and exceptions]
DataRoom --> Release[Release decision]The data room should answer five questions:
- What system did we intend to build?
- What system did we actually assemble?
- What does it do reliably?
- What does it fail to do?
- Who approved the release and under what limitations?
43.6 Limited availability
Limited availability is the safest way to learn from real users without pretending the platform is mature.
View diagram source
sequenceDiagram
participant User as Approved user
participant Portal as Platform portal
participant Triage as Admission review
participant QPU as New QPU
participant Ops as Operations
User->>Portal: submit candidate workload
Portal->>Triage: attach workload metadata
Triage->>Ops: request readiness check
Ops-->>Triage: approve window and constraints
Triage->>QPU: schedule workload
QPU-->>Portal: result plus provenance
Portal-->>User: return qualified resultLimited availability should expose explicit restrictions:
- supported circuit families,
- maximum depth or duration,
- supported dynamic features,
- job volume limits,
- expected queue behavior,
- data-retention terms,
- known device limitations,
- escalation process.
43.7 Commissioning anti-patterns
| Anti-pattern | Consequence |
|---|---|
| qubit-count release | users mistake width for useful capacity |
| benchmark-only release | operational failures appear after launch |
| undocumented waivers | future regressions become untraceable |
| heroic calibration | platform works only when one expert is present |
| no shadow traffic | scheduler and data paths are untested |
| missing rollback | bad releases become political instead of technical |
View diagram source
flowchart TD
Anti[Commissioning anti-pattern] --> Count[Qubit-count marketing]
Anti --> Bench[Benchmark-only evidence]
Anti --> Hero[Heroic specialist workflow]
Anti --> Waiver[Untracked waivers]
Anti --> Rollback[No rollback path]
Count --> Bad[False confidence]
Bench --> Bad
Hero --> Bad
Waiver --> Bad
Rollback --> Bad43.8 Operating rule
A quantum system is commissioned when the organization can state:
This is the system, these are its validated operating envelopes, these are the workloads we will accept, these are the artifacts we preserve, these are the failure modes we know, and this is how we roll back.
Anything less is still bring-up.