Skip to main content

​
End-to-end inference platform execution plan

InferCrane owns the operational lifecycle of inference, regardless of where model execution happens:
This plan is dependency ordered. A later layer may not become authoritative over an earlier one. Provider, runtime, gateway, evaluator, and sandbox implementations remain replaceable adapters.

​
Product rules

  • Do not build a second inference engine, GPU scheduler, router, cloud provisioner, sandbox runtime, or workflow engine.
  • Do not use an LLM for routing, promotion, budget, or rollback decisions.
  • Do not turn simulated provider behavior into a real-provider claim.
  • Do not publish managed compute before billing and abuse risk have explicit owners.
  • Prefer a five-minute existing-endpoint connection over forcing migration.
  • Keep sandbox and training integrations behind the boundaries in ADR 0032; they are not immediate-launch capabilities.

​
Success measures

  • First successful proxied request in under five minutes for an existing endpoint.
  • One application model alias remains stable while its serving plan changes.
  • Every external transmission has consent, provider identity, and a pre-authorized hard budget.
  • Every release decision can be reconstructed from immutable evidence.
  • Hosted-console use does not weaken local/self-hosted authentication or authorization.