Teams working near emerging technology, whether quantum hardware, simulators or experimental research clusters, tend to run unusual infrastructure with very few people watching it. That is precisely where an alert pipeline either earns its keep or becomes a liability. This is an audit checklist for assessing whether an operations environment is ready for AIOps-style alert handling, written for engineers who prefer to tick boxes with evidence attached. Nothing here claims that quantum methods help with alerting; the parallels to optimisation and machine learning are noted only where they clarify the engineering.
The reference throughout is the Ops Atlas, whose entry on AIOps, an AIOps reference with scope, maturity gates and security notes, defines the discipline as artificial intelligence applied to operations rather than operations of AI systems. It also makes the point that suits an audit: the first four practices involve no machine learning, so you can verify them with a spreadsheet.
Section A: Preconditions (stop here if any item fails)
- A1. Working alert rules exist. Evidence: a rule inventory where each rule has fired and been acted on at least once in the review period.
- A2. Every service has a named owner. Evidence: a table of services with a person or team and a contact route.
- A3. A dependency list exists. Evidence: a document or graph that says which components rely on which, with a last-reviewed date.
- A4. A change log exists. Evidence: a record of deployments and configuration edits, time-stamped and queryable by service.
The Atlas treats these four as prerequisites, and for good reason: the Google SRE book it cites links about 70 percent of outages to alterations of a live system, and a team with no change record has little for correlation to chew on.
Section B: Label contract
- B1. A mandatory label set is documented. The Atlas example covers the service, the owning team, severity, the remediation, the affected cluster, the evaluating Prometheus server and a runbook URL.
- B2. No label carries two meanings. Test: confirm the tag for the affected cluster is separate from the tag for the evaluating server. The Atlas documents a runbook where one overloaded tag let an inhibition rule mute a real outage.
- B3. Compliance is measured automatically. Evidence: a lint step on the rule repository plus a compliance measurement taken from the live alert stream.
- B4. Compliance has reached the gate. The Atlas sets 95 percent as the exit criterion for its first maturity level.
Section C: Deduplication and grouping
- C1. Fingerprints exclude volatile values. Test: restart a workload and confirm the new alert collapses into the old one instead of appearing as a stranger.
- C2. Grouping keys are chosen deliberately. Evidence: documented keys per service and an example of a grouped notification.
- C3. Every suppression rule names mandatory labels. The Atlas notes a quirk: if all listed matching labels are absent from both the source and target alert, the router applies suppression regardless. Test with an alert pair carrying none of them.
- C4. Suppressed alerts are visible. The Atlas runbook starts with a command-line query that lists inhibited alerts. Evidence: that query is saved in the on-call guide and a dashboard counts active suppressions.
- C5. Unreachable-cluster alerts require two vantage points. A single observer losing its own network can otherwise hide everything.
Section D: Scoring
- D1. A labelled incident set exists. The Atlas suggests about a hundred before a learned grouping goes into shadow trial.
- D2. Pairwise precision, recall and F1 are computed. Not just compression ratio, which the Atlas calls a vanity figure: its lab saw a simple time window shrink alert counts severalfold while merging unrelated incidents.
- D3. Topology changes are re-scored. In the Atlas experiment, removing one dependency edge cost a great deal of recall, so evidence should show the map is revalidated after architecture changes.
- D4. Rule quality attributes are tracked. The SRE Workbook lists four attributes to record per rule: precision, recall, detection speed and reset speed.
Readers familiar with machine-learning evaluation will find this section comfortable ground; our overview of the ways quantum computing could reshape learning algorithms explains why evaluation discipline matters in any learning system, quantum or not. Likewise, the article on quantum algorithms for hard optimisation problems frames grouping and ranking tasks as optimisation, which is a useful mental model but not a claim about current tooling.
Section E: Security trust boundaries
This section matters most in research environments, where operators often relax controls for convenience.
- E1. The alert router API is not open to any web origin. The Atlas observes that its default API answers cross-origin requests from anywhere, so one request could create a silence that hides an outage. Evidence: restricted origins, authentication and logged silences with automatic expiry.
- E2. Ingest webhooks are authenticated.
- E3. Event-driven automation sources are not exposed. The Atlas notes that the alert-router source in a widely used automation collection (version 2.13.0) accepts unauthenticated HTTP, so it should listen only on a loopback address.
- E4. The remediation runner is treated as a high-value target. It holds production credentials. Evidence: scoped accounts, short-lived secrets and audit logs.
- E5. Remediation executes under the account that owns the files involved, never root. The Atlas quotes the automation tool’s documentation: you cannot limit privilege escalation to certain commands.
- E6. Alert text is treated as untrusted input. Anything a language model reads may have been written by an attacker.
Section F: Automation and lifecycle
- F1. Fixes are restricted to an approved list, can be undone and are throttled. The Atlas lab example limits a log-rotation job per machine per day and across the whole estate per hour, then verifies the outcome and pages a human if verification fails.
- F2. A policy lifecycle is in use. Proposed, shadow, canary, active, review and retired, with named rollback triggers.
- F3. Components are portable. Rules, labels and evaluation sets can move between products. Cloud anomaly-detection services have been shut, retired or closed to newcomers across 2025 and 2026, so portability is a practical test rather than a theoretical one.
- F4. Review cadence is in the calendar. Weekly noise review, monthly scorecard, quarterly platform review.
Notes for auditors working with unusual hardware
Research clusters and lab equipment produce alerts that standard rule sets do not anticipate: calibration drifts, scheduled maintenance windows, experimental jobs that deliberately push a system to its limits. Record these in the same calendar that you use for holidays, since the Atlas lab showed that undeclared calendar events were responsible for nearly all the false alarms from a seasonal baseline. Ask the scientists or engineers who run the equipment which alerts they already ignore, and treat each answer as a candidate for a rule fix or a group. Finally, check that the audit itself does not create risk: read-only access is enough for every item in Sections A to D, and Section E tests should be run against a copy of the alert router, not production.
Scoring the audit
Give one point per item with attached evidence and zero for assertions without proof. A fail on any Section A item means the programme should not proceed to automation. A result of fewer than three points in Section E should block any remediation runner from production regardless of the total. Re-run the audit quarterly and keep the previous version for comparison; the changes between audits are usually more informative than the score itself.
If you adopt only one finding from this checklist, make it the order of work. According to the Atlas, plain steps lead and model-driven ones follow. It adds that leaderboard results for language-model agents, with diagnosis scores of 56.2 percent, cover narrower tasks than live resolution and do not justify letting them act unsupervised.