Skip to main content

​
Deterministic explanations

InferCrane explanations are projections of persisted state, events, decisions, policies, and measurements. They do not call an LLM and do not infer facts that were not recorded.
The general explanation reports observed deployment state, unhealthy replica evidence, and the active durable operation that is blocking convergence, including its persisted kind, status, progress, message, and error code. Scaling explanations reproduce the latest persisted autoscaling action, reason, signal snapshot, and evaluation timestamp. Rollout explanations reproduce the Release Guard evaluation ID, active and candidate revisions, decision, reason codes, metric snapshot, policy snapshot, and timestamp. Cold-start explanations use the provider-worker observation and TTFT evidence described in Cold-start intelligence. Every form supports --output json. Repeating an explanation against unchanged persisted state produces the same explanation code and evidence. If no evaluation exists, InferCrane says so instead of inventing a cause.