Cross-Provider Migration and Portability Drills

Portability is not a promise until it has been drilled. A workload that compiles for several providers may still depend on hidden assumptions: measurement ordering, basis gates, dynamic-control support, mitigation defaults, queue behavior, shot limits, data schemas, or regional policy.

This chapter turns provider portability into an operational practice, not an architectural aspiration.

flowchart LR
    Workload[Workload package] --> Abstract[Abstract contract]
    Abstract --> ProviderA[Provider A]
    Abstract --> ProviderB[Provider B]
    Abstract --> ProviderC[Provider C]
    ProviderA --> Compare[Compare outputs and evidence]
    ProviderB --> Compare
    ProviderC --> Compare

Portability dimensions

A portable quantum workload requires more than source-code translation.

mindmap
  root((Portability))
    Semantics
    Compiler
    Hardware
    Runtime
    Evidence
    Governance
    Economics

The portability contract

A portability contract defines what must remain invariant and what may vary.

classDiagram
    class PortabilityContract {
      workload_id
      invariant_semantics
      acceptable_transformations
      required_evidence
      comparable_metrics
      disallowed_provider_features
      exit_thresholds
    }
    class ProviderAdapter {
      provider
      target_mapping
      compiler_profile
      runtime_profile
      evidence_mapper
    }
    PortabilityContract --> ProviderAdapter

Invariants may include the Hamiltonian, objective function, logical circuit, random seed policy, required confidence level and precision, or claim boundary. Variables may include the observed confidence interval, physical layout, shot allocation, mitigation method, and queue timing.

Migration drill lifecycle

flowchart TB
    Select[Select workload] --> Package[Package evidence and contract]
    Package --> Translate[Translate/adapt]
    Translate --> DryRun[Simulator or dry run]
    DryRun --> Execute[Execute on alternate provider]
    Execute --> Compare[Compare result and evidence]
    Compare --> Report[Migration report]
    Report --> Fix[Fix portability gaps]

A drill should produce a report even if migration fails. Failure identifies coupling that architecture diagrams usually hide.

Comparison is not equality

Cross-provider results rarely match exactly. The question is whether differences remain inside the claim boundary.

flowchart LR
    ResultA[Provider A distribution] --> Distance[Distance and interval analysis]
    ResultB[Provider B distribution] --> Distance
    Distance --> Boundary{Inside claim boundary?}
    Boundary -- yes --> Portable[Portable enough]
    Boundary -- no --> Gap[Portability gap]

Comparison techniques may include distribution distance with uncertainty bounds, an interval for the difference between observables, benchmark-relative scoring, or task-specific success probability. Predeclare practical equivalence margins and the comparison method. Overlap between two separate confidence intervals, or failure to detect a difference, does not establish equivalence.

Evidence mapping

Evidence mapping is often harder than execution. Each provider exposes different metadata and runtime semantics.

flowchart TB
    ProviderEvidence[Provider evidence] --> Mapper[Evidence mapper]
    Mapper --> Canonical[Canonical evidence schema]
    Canonical --> Validator[Schema validator]
    Validator --> Archive[Evidence archive]

Exit strategy drills

An exit strategy should be tested before it is needed. The test does not need to migrate every workload. It should migrate representative workload classes.

flowchart LR
    Dependency[Provider dependency] --> Drill[Exit drill]
    Drill --> Gaps[Gap register]
    Gaps --> Remediation[Remediation backlog]
    Remediation --> Retest[Retest]
    Retest --> Confidence[Exit confidence]

Exit confidence depends on workload coverage, evidence portability, contract flexibility, staff familiarity, and cost impact.

Operating rule

Do not describe a workload as portable until a migration drill has produced comparable results, canonical evidence, and a gap register.

Additional technical sources: [R269], [R271].