A support-ticket classifier's 3,000-token policy document is billed as fresh input on every single call, even though it hasn't changed in weeks. What actually fixes this?

Answer: B — Move it into a cached system block, checkpointed before any per-ticket content

Nothing about the policy needs to change — it's uncached, not too long. A checkpoint after the stable content bills it once and reuses it. (A) throws away real content; (C) trades quality for a problem that isn't quality-related; (D) solves a latency-tolerance issue that was never the actual waste.