High-Assurance Validation and Audit — Operating Quantum Computers

Some quantum results will remain exploratory. Others will influence funding, product claims, customer decisions, cryptographic migration, materials strategy, or safety-relevant engineering. High-assurance workflows are the bridge between scientific exploration and accountable decision-making.

High assurance does not mean the result is certainly correct. It means the uncertainty, assumptions, controls, evidence, and review trail are strong enough for the intended decision.

65.1 Assurance levels

Diagram loads as you read

65.1 Assurance levels · Figure 1

View diagram source

flowchart TB
    Result[Quantum result] --> Level{Assurance level}
    Level --> L0[Exploratory]
    Level --> L1[Internal decision]
    Level --> L2[External claim]
    Level --> L3[Regulated or safety-relevant]

Each level should impose a different review burden.

Level	Use	Required control
L0	exploration	owner, manifest, basic reproducibility
L1	internal decision	baseline, uncertainty, evidence package
L2	external claim	independent review, locked artifacts, approved language
L3	regulated/safety-relevant	formal audit trail, segregation of duties, retention policy

65.2 Claim decomposition

Diagram loads as you read

65.2 Claim decomposition · Figure 2

View diagram source

flowchart LR
    Claim[Claim] --> Scientific[Scientific assertion]
    Claim --> Operational[Operational assertion]
    Claim --> Economic[Economic assertion]
    Claim --> Comparative[Comparative assertion]

A statement such as “this quantum workflow improves optimization quality” might contain several claims: the formulation is valid, the quantum result is statistically significant, the baseline is fair, and the cost is acceptable. Review each separately.

65.3 Evidence checklist

Diagram loads as you read

65.3 Evidence checklist · Figure 3

View diagram source

flowchart TB
    Evidence[Evidence package] --> Inputs[Inputs]
    Evidence --> Execution[Execution]
    Evidence --> Statistics[Statistics]
    Evidence --> Baselines[Baselines]
    Evidence --> Review[Review trail]

A high-assurance evidence package should include:

problem statement
source and artifact hashes
target profile and calibration snapshot
circuit or workload contract
raw and processed results
statistical method
mitigation policy
baseline comparison
sensitivity analysis
reviewer notes
approved claim language

65.4 Independent reproduction

Diagram loads as you read

65.4 Independent reproduction · Figure 4

View diagram source

sequenceDiagram
    participant Author as Author team
    participant Store as Evidence store
    participant Reviewer as Independent reviewer
    participant Runtime as Runtime
    Author->>Store: publish locked evidence package
    Reviewer->>Store: retrieve manifest and artifacts
    Reviewer->>Runtime: reproduce selected runs
    Runtime-->>Reviewer: reproduction results
    Reviewer->>Store: attach review decision

Reproduction does not always require repeating every shot. It should repeat enough of the chain to validate the claim's sensitivity.

65.5 Audit trail design

Diagram loads as you read

65.5 Audit trail design · Figure 5

View diagram source

erDiagram
    CLAIM ||--o{ REVIEW_DECISION : has
    CLAIM ||--o{ EVIDENCE_PACKAGE : supported_by
    REVIEW_DECISION }o--|| REVIEWER : made_by
    EVIDENCE_PACKAGE ||--o{ ARTIFACT_HASH : contains
    EVIDENCE_PACKAGE ||--o{ POLICY_CHECK : passed

    REVIEW_DECISION {
        string decision_id
        string status
        datetime decided_at
        string rationale
    }

The audit trail must answer who approved the claim, what they saw, which policies passed, and whether any artifacts changed after approval.

65.6 Red-team review

Diagram loads as you read

65.6 Red-team review · Figure 6

View diagram source

flowchart TB
    Claim[Claim] --> RedTeam[Red-team review]
    RedTeam --> Baseline[Baseline challenge]
    RedTeam --> Stats[Statistics challenge]
    RedTeam --> Noise[Noise/mitigation challenge]
    RedTeam --> Repro[Reproducibility challenge]
    RedTeam --> Language[Language challenge]

High-value claims deserve adversarial review. The reviewer should ask what would make the claim false.

65.7 Statistical sign-off

Diagram loads as you read

65.7 Statistical sign-off · Figure 7

View diagram source

flowchart LR
    Result[Measured result] --> Uncertainty[Uncertainty model]
    Uncertainty --> Sensitivity[Sensitivity analysis]
    Sensitivity --> DecisionThreshold[Decision threshold]
    DecisionThreshold --> Signoff{Sign off?}

The sign-off should define:

minimum effect size
confidence level
multiple-comparison handling
stopping rule
uncertainty propagation
mitigation-bias caveats

65.8 Claim language control

Diagram loads as you read

65.8 Claim language control · Figure 8

View diagram source

stateDiagram-v2
    [*] --> DraftClaim
    DraftClaim --> TechnicalReview
    TechnicalReview --> LegalReview: external claim
    TechnicalReview --> Revision: needs change
    LegalReview --> Approved
    Approved --> Published
    Published --> Retracted: evidence invalidated

Quantum claims are easy to overstate. Separate result review from language approval when external stakeholders are involved.

65.9 Continuous audit

Diagram loads as you read

65.9 Continuous audit · Figure 9

View diagram source

flowchart LR
    Approved[Approved claim] --> Monitor[Monitor dependencies]
    Monitor --> Target[Target degradation]
    Monitor --> Source[Source superseded]
    Monitor --> Baseline[Baseline improved]
    Monitor --> Standard[Standard changed]
    Target --> Reassess[Reassess claim]
    Source --> Reassess
    Baseline --> Reassess
    Standard --> Reassess

A claim can become stale even if the original evidence was sound.

65.10 Operating rule

Assurance is proportional to consequence.

Diagram loads as you read

65.10 Operating rule · Figure 10

View diagram source

flowchart LR
    Consequence[Consequence of being wrong] --> Assurance[Required assurance]
    Assurance --> Controls[Controls]
    Controls --> Trust[Trust]

Exploration should remain fast. Decision-grade claims should be slow enough to be trustworthy.

Additional technical sources: [R267].