Foundation models now underpin production software, yet the infrastructure that trains and serves them is brittle: long training runs fail silently, serving clusters saturate under bursty demand, and silent model drift degrades output quality without tripping conventional alarms. This paper applies reliability engineering to foundation model infrastructure. We present the Keystone Reliability Framework, which maps four problem domains to concrete practices, mechanisms, and measurable outcomes, and we evaluate the framework through fault injection on a representative training and serving testbed. Across the injected fault scenarios, checkpointing, error-budget-driven admission control, and robustness guards reduced mean recovery time and held output quality within target service levels relative to an unguarded baseline. The results are illustrative rather than production-measured, but they show that classic reliability principles transfer to foundation model systems when adapted to their distinctive failure modes.