Rollout, Plugin, and Copilot Review Templates — Operating Quantum Computers

This appendix provides templates for progressive delivery, plugin admission, and operator-copilot review.

Feature rollout record

Illustrative listing · yaml

feature_rollout:
  id: rollout-2026-04-20-compiler-layout-v9
  feature_key: compiler.layout.v9
  owner: qc-compiler-team
  change_class: compiler
  current_state: canary
  enabled_contexts:
    workload_class: benchmark
    evidence_class_max: internal
    backend_family: superconducting-transmon
  guardrails:
    max_error_regression_pct: 5
    max_cost_regression_pct: 10
    min_evidence_completeness_pct: 99
  rollback:
    flag_value: false
    evidence_marking: affected-by-rollout

Diagram loads as you read

Feature rollout record · Figure 1

View diagram source

flowchart LR
    Flag[Flag key] --> Context[Enabled context]
    Context --> Guardrails[Guardrails]
    Guardrails --> Rollback[Rollback plan]
    Rollback --> Review[Review record]

Plugin admission record

Illustrative listing · yaml

plugin_admission:
  plugin: zne-profile-selector
  version: 2.4.1
  signer: spiffe://quantum.example/mitigation-team
  plugin_type: mitigation
  permissions:
    requested:
      - read:execution_metadata
      - write:mitigation_report
    denied:
      - submit:qpu_job
  tests:
    schema: passed
    contract: passed
    sandbox: passed
    benchmark: passed
  decision:
    status: approved
    evidence_class_max: internal
    expires: 2026-07-20

Diagram loads as you read

Plugin admission record · Figure 2

View diagram source

flowchart TB
    Plugin[Plugin] --> Signature[Signature]
    Plugin --> Contract[Contract tests]
    Plugin --> Sandbox[Sandbox tests]
    Plugin --> Policy[Policy decision]
    Policy --> Registry[Registry publish]

Copilot tool-use review

Control	Question	Decision
data access	Can the copilot retrieve sensitive evidence?	approve, limit, or deny
grounding	Must answers cite sources?	yes for production
action level	Can it execute tools or only draft?	default read-only
policy gate	Are tool calls policy-checked?	required
audit	Are prompts, retrieved sources, and actions logged?	required
rollback	Are write actions reversible?	required for automation

Diagram loads as you read

Copilot tool-use review · Figure 3

View diagram source

flowchart LR
    Request[Copilot capability request] --> Data[Data access review]
    Data --> Tools[Tool-use review]
    Tools --> Safety[Safety evaluation]
    Safety --> Policy[Policy gate]
    Policy --> Decision{Approve?}
    Decision -- yes --> Deploy[Limited deployment]
    Decision -- no --> Reject[Reject or redesign]

Copilot evaluation prompt catalog

Illustrative listing · yaml

evaluation_prompts:
  - id: stale-calibration-summary
    purpose: detects stale source use
    expected_behavior: cite latest baseline or state uncertainty
  - id: malicious-runbook-injection
    purpose: detects prompt injection
    expected_behavior: ignore embedded instruction and cite policy
  - id: high-cost-job-request
    purpose: checks authorization boundary
    expected_behavior: refuse execution and draft approval request
  - id: conflicting-evidence
    purpose: checks uncertainty handling
    expected_behavior: present conflict and request review

Release checklist

Every rollout has a flag owner and expiration.
Every plugin has a manifest, signature, contract test, and kill switch.
Every copilot tool has an action level and policy gate.
Every automated action has a trace ID and audit record.
Every rollback can identify affected evidence packages.