Overview
Connecting…Service map · live traffic
Business health · order flow5-minute window
ConfirmedFailedPending review
WebLogic soa_server2
Stuck threadsDatasource activeJMS depth ÷10
Latest remediator run
passfixedflaggedblockedone square per control, in order
Approval queue
Inject a faulteach one makes a real change to the lab · click again to clear
AI controls this session
Kubernetes eventsnewest first
Incidents
Remediator runs · 3 checks each
| Check | Model said | AI controls | Result |
|---|
AI scorecardthis session
The ten controls
| # | Control | What it does |
|---|---|---|
| 1 | Collect | Read-only snapshot: deployments, pods, HPA, events, metrics, logs |
| 2 | Redact | Masks passwords, JDBC credentials, tokens and OCIDs before the model sees them |
| 3 | Fence | Logs go inside <untrusted> tags; instruction-like text is flagged |
| 4 | Model | One focused prompt per check, temperature 0, fixed JSON schema |
| 5 | Schema | Required fields; the action must be a pipeline ID or "none", never a command |
| 6 | Evidence | Every pod, HPA, metric or log the model cites must exist right now |
| 7 | Guardrail | A deterministic rule must agree, e.g. HPA at max AND CPU above target for 30 s |
| 8 | Allow-list | Only rollback-image, raise-hpa-max, restart-managed-server on known targets |
| 9 | Dedup | One incident per problem; ImagePull takes priority over business alerts |
| 10 | Approval & audit | Nothing changes until a person (or an explicit policy) approves; then execute, verify, log |
Evaluation suite · does the AI layer earn its place?
Injects eight faults one after another against the live lab, runs the remediator and scores the outcome. Run it once with AI controls ON and once OFF, or once per model, and compare the runs below. Takes about 10 to 15 minutes. Incidents raised during the suite are closed automatically.
| Scenario | Expected | Result | Detail | Time to detect |
|---|
Previous evaluation runscompare controls ON vs OFF, or one model against another
| Finished | AI controls | Model | Passed | Scenarios |
|---|
Model callslast 15
Audit log
every decision, newest firstExport session (JSON)
Model providerRuns in this browser tab; keys are never saved
Guardrail thresholdsapplied on the next run
Approval policywhich fixes may skip human approval
Everything needs approval by default. Ticking an action lets it run as soon as every other control passes. A typical policy auto-approves an image rollback and keeps restarts and capacity changes manual.
Notifications
Slack and Teams notifications are sent by the laptop lab only. In this demo, each notification is written to the audit log instead.
Appearance