Skip to main content

​
Product qualification

InferCrane has one top-level qualification command. It composes the maintained unit, race, PostgreSQL, Docker, Kind, black-box journey, provider-contract, SDK, Terraform, security, documentation, supply-chain, and real-provider suites. It does not replace those suites or hide their logs. Overlapping release work runs once: the black-box gate runs user journeys, while the supply-chain gate owns archive, checksum, SBOM, and Homebrew verification.
The command is resumable. A passed gate is reused only for the exact Git commit. The worktree must be clean, so uncommitted changes can never inherit release evidence from an earlier source tree. Evidence is written to:
Inspect the current matrix without running tests:
Longer fuzz, soak, and fleet simulation run separately so pull-request feedback remains fast:
See Reliability test environment for the evidence ladder, provider simulation boundaries, version matrix, and scheduled workflow. The verdict vocabulary is intentionally strict:

​
Real RunPod qualification

Paid provider mutation always requires an explicit approval flag. Reuse the run ID after a shell disconnect; durable operations and stage evidence allow the workflow to resume.
The command records four independent gates: serverless lifecycle, serverless faults, elastic lifecycle, and elastic faults. A provider allocation failure in one gate cannot erase successful evidence from another. Together they cover vLLM protocols, AIPerf, autoscaling, Release Guard, streaming cancellation, lost-response adoption, disruption, deletion, and direct zero-inventory checks. A provider allocation or image-transfer failure remains a failed suite even when guarded cleanup succeeds.

​
AWS, GCP, and Kubernetes GPU qualification

The portable qualification runner needs an isolated provider environment, immutable workload specs, and one API key shared by the control plane and workers. See provider setup before running any paid gate.
The compatibility-prefixed INFERCRANE_V1_* names do not select an old release. A gate remains REAL_INFRA_REQUIRED until its exact-commit paid run passes; the existence of the guarded runner is not qualification evidence.

​
What the system proves

Local qualification proves control-plane behavior against deterministic dependencies, a real PostgreSQL process, Docker process failures, and a real Kind API. It cannot prove GPU availability, provider eventual consistency, customer networking, runtime performance, public package delivery, or human usability. Those boundaries remain separate rows in report.json so automation can enforce release policy without turning simulated evidence into a production claim.