Free-tier AI prototypes fail at runtime, not the model call, and need a readiness checklist

A DEV Community article argues that hosted AI prototypes on free tiers tend to break from server-side issues rather than model calls. It contends that local disk may be ephemeral, instances can sleep or be recycled, addresses can change between deploys, and outbound model calls may be blocked or rate-limited. The author proposes a runtime-readiness checklist where each claim needs a named evidence artifact, and any claim lacking recent evidence is treated as blocked rather than passed. The checklist covers cold starts, state durability, idempotent writes, payload limits, egress, header leakage, and rollback, and includes a stdlib-only Python probe that writes one evidence file per gate.