Skip to main content

​
FinOps and advisory Autopilot

InferCrane treats cost as evidence, not a guess. A FinOps report accepts only cost metadata carrying a source and observation time. Missing cost remains unavailable, currencies are never silently converted, and avoidable spend remains unknown until utilization and a qualified alternative directly support it. For Kubernetes workloads already measured by OpenCost, import one or more exact allocation keys:
The collector uses OpenCost’s standard /allocation API. It imports only allocation keys named by --allocation; InferCrane never treats whole-cluster cost as one deployment implicitly. OpenCost’s window total becomes an average hourly rate over that same persisted window. The result is classified as provider-reported because it is derived from resource allocation and pricing inputs rather than a settled invoice. Use --file allocation.json for an offline response.
The report persists its exact evidence, evaluation window, disclosure set, and canonical input digest. The current release reports only fresh sourced hourly cost that InferCrane can bind to the selected workload, immutable revision, and window. Cost per request, per-token unit economics, idle-capacity attribution, and verified savings remain unavailable unless request coverage, billing units, currency, and a comparable post-change window are all grounded. The product never silently mixes currencies or turns a provider list price into realized savings.

​
Advisory Autopilot

Autopilot consumes the deterministic SLO recommendation already persisted for a deployment:
plan creates an immutable evidence snapshot. approve records the authenticated human decision. Neither command changes a deployment, provisions capacity, or promotes a revision. The next production step is an ordinary explicit candidate rollout followed by Replay, benchmark, and Release Guard.
Approval is not execution. InferCrane intentionally has no autonomous production mutation path.