Cache the Plan, Not the Answer
I measured what data drift does to LLM answer caches: 616,200 scored queries across two model families. Semantic caches rot faster than the data, invalidation alone creates a new failure mode, and the fix is to cache the plan, not the answer.