Autonomous Calibration and AI Operations

Automation is unavoidable. State-of-the-art quantum systems expose too many parameters, too many correlations, and too much temporal variability for purely manual operation. But autonomous operation must be bounded. The goal is not to “let AI run the quantum computer.” The goal is to turn repetitive decisions into controlled policies with evidence, guardrails, rollback, and human accountability.

45.1 Automation ladder

flowchart LR
    Manual[Manual specialist action] --> Assisted[AI-assisted recommendation]
    Assisted --> Supervised[Human-approved automation]
    Supervised --> Bounded[Bounded autonomous control]
    Bounded --> Closed[Closed-loop operation]

Most organizations should spend a long time in the middle: recommendations, assisted analysis, and supervised automation. Closed-loop autonomy is appropriate only when blast radius is constrained and rollback is tested.

NVIDIA’s Ising family includes AI models for calibration and decoding. These tools still require validated operating limits and governance. The Qiskit Experiments calibration-management APIs linked in R84 are historical and have been removed; implementations need currently supported controller and calibration interfaces. [R83] [R84]

45.2 Control-loop anatomy

flowchart TB
    Sensors[Telemetry and experiment results] --> Features[Feature extraction]
    Features --> Policy[Policy or model]
    Policy --> Proposal[Proposed action]
    Proposal --> Guardrails[Safety and validity guardrails]
    Guardrails --> Approval{Approval mode}
    Approval -- manual --> Human[Human reviewer]
    Approval -- bounded auto --> Actuator[Calibration or routing actuator]
    Human --> Actuator
    Actuator --> Observe[Observe outcome]
    Observe --> Sensors

The model proposes. The platform constrains. The audit trail records.

45.3 Suitable targets for AI assistance

flowchart TD
    AI[AI operations use] --> Recommend[Recommend]
    AI --> Detect[Detect]
    AI --> Rank[Rank]
    AI --> Summarize[Summarize]
    AI --> Control[Control]
    Control --> Guarded[Only with guardrails]

45.4 Guardrail design

Autonomous operations need hard limits.

autonomy_guardrail:
  control_surface: calibration_parameter | scheduler_policy | decoder_version | analysis_pipeline
  permitted_action: string
  max_delta_per_action: string
  max_actions_per_window: integer
  rollback_artifact: string
  human_approval_required_when:
    - impact exceeds review threshold but remains within hard limits
    - uncertainty exceeds autonomous-action threshold
  block_when:
    - safety invariant violated
    - action exceeds hard physical limits

flowchart LR
    Proposal[Model proposal] --> Bounds{Within bounds?}
    Bounds -- no --> Block[Block and alert]
    Bounds -- yes --> Confidence{Confidence sufficient?}
    Confidence -- no --> Review[Human review]
    Confidence -- yes --> Canary[Canary action]
    Canary --> Evaluate{Improved?}
    Evaluate -- no --> Rollback[Rollback]
    Evaluate -- yes --> Promote[Promote]

45.5 Model promotion

AI models used in operations must pass the same release discipline as compilers, calibrations, and runtime images.

sequenceDiagram
    participant Dev as Model developer
    participant Eval as Offline evaluation
    participant Shadow as Shadow mode
    participant Canary as Canary deployment
    participant Prod as Production policy
    Dev->>Eval: train candidate
    Eval-->>Dev: metrics and failure cases
    Dev->>Shadow: deploy read-only
    Shadow-->>Canary: approve limited action
    Canary-->>Prod: promote if gates pass

Promotion criteria should include:

• offline performance,
• out-of-distribution detection,
• false-positive and false-negative cost,
• traceability of training data,
• reproducibility of model builds,
• interaction with human review,
• rollback time,
• incident playbook.

NIST’s AI RMF is not quantum-specific, but its emphasis on managing AI risks across design, development, deployment, and use maps directly to autonomous quantum operations. [R87]

45.6 Human accountability

Automation must not erase accountability.

flowchart TB
    Action[Automated action] --> Owner[Named service owner]
    Action --> Policy[Versioned policy]
    Action --> Evidence[Evidence bundle]
    Action --> Rollback[Rollback path]
    Action --> Review[Periodic review]

For each autonomous action, the system should know:

1. which policy authorized the action,
2. which model version proposed it,
3. which telemetry supported it,
4. which guardrails accepted it,
5. which owner is accountable,
6. how to reverse it.

45.7 Failure modes

flowchart TD
    Model[AI operations model] --> Proxy[Optimizes proxy]
    Proxy --> Drift[Physical state drifts]
    Drift --> BadAction[Bad action looks plausible]
    BadAction --> Guardrail{Guardrail catches?}
    Guardrail -- yes --> Learn[Record failure case]
    Guardrail -- no --> Incident[Operational incident]

45.8 Operating rule

Autonomous quantum operations should be judged by reversibility, observability, and bounded blast radius. A model that improves calibration speed but cannot explain, constrain, or reverse its actions is not production automation. It is an incident waiting for a window.