output "topic_ocid" { description = "The topic every alarm publishes to. Subscribe other endpoints here." value = oci_ons_notification_topic.ai.id } output "verify" { description = "How to prove each alarm fires, that the event names match, and whether Audit records model calls." value = <<-EOT Confirm each subscription email first. Audit usually records a call within 15 minutes, so allow that long for each email. In a test compartment, create a Generative AI API key in the console, then delete it # ai-api-key-changed emails, naming you Update a test alarm's display name # ai-logging-weakened follows If you run a dedicated AI cluster, edit a test endpoint's description and set it back # ai-endpoint-changed follows ai-hosting-created has no cheap test: a hosting cluster commits to 744 unit-hours. Its connector is built like the other three. If one stays quiet, search the tenancy's _Audit_Include_Subcompartment for the call and read data.eventName: the connector filters on that exact name. Make one on-demand Chat call in a test compartment, then run the saved search ai-inference-calls over the last hour. Nothing listed means Audit does not record model calls in ${var.region}. Each alarm's body names its saved search in Logging Search. EOT }