Deployment and governance
Evidence, gates, rollback
The Problem
What production adds
Production introduces change management, access control, budget caps, incident response, and contractual SLAs. A successful lab demo answers scientific questions; a pilot answers operational ones.
This closing chapter synthesises cross-cutting requirements from the technical chapters: encoding discipline, baselines, and evidence you can defend under review.
The Challenge
Integration risk dominates asymptotic complexity
The Challenge
Where programmes stall
Teams discover late that batching policies, data residency rules, or latency budgets invalidate a cute circuit. The mitigation is to run the hybrid stack inside the same CI harness you use for classical models, with added hooks for shot noise and calibration drift.
Programme risks to name explicitly
The Solution
Promotion gates written like contract clauses
The Solution
What a gate contains
Each gate states a measurement, a threshold, an owner, and a remediation action when the threshold fails. “Improve the quantum model” is not a remediation; “disable layer X and re-run parity suite” is.
Minimum gate set
- Statistical parity on locked validation data within agreed tolerance.
- Runtime p95 and p99 under realistic batching and region placement.
- Reproducibility: dataset hash, dependency lockfile, seed list, shot schedule.
Implementation
YAML-style charter + CI hooks
Implementation
Encode gates in version control so diffs are reviewable. The pattern below is representative, not exhaustive.
Pair with automated evaluation jobs triggered on each merge request touching quantum code paths.
pilot:
owner: platform@example.com
success_metrics:
- name: val_accuracy_gap_vs_classical
max: 0.005
- name: latency_p95_ms
max: 120
rollback:
traffic: blue_green_to_classical_v3
approver: sre@example.comSummary
Ship discipline, not mystique
Summary
Closing reminder
The workshop materials deliberately run on simulators so every stakeholder can reproduce the chain of thought. When you move toward specialised hardware, keep the artefacts—QUBO definitions, logs, classical baselines—identical in spirit so comparisons remain legible.