The attractive story is simple: keep expensive judgment for consequential work and push routine human-facing summaries onto cheaper or local lanes. That is good cost discipline. It is also incomplete if the new lane has never swallowed the real job packet.
The failure mode is quiet until it is not. A local model can exist, route correctly, and still be unusable for a specific delivery job because the runtime context, tool-call behavior, or prompt shape does not match the contract the job actually needs. “Local-first” is a direction, not proof of readiness.
The portable rule is to flight-check the lane before trusting it: run the exact packet through the actual runner, verify the context window and tool contract, prove fallback behavior, and inspect the human-facing receipt. Cheap is only cheaper after it works.