Why it matters
Most token blowups are not decisions, they are defects. A context document added to every request, or an unintended multi-turn loop, can raise consumption tenfold overnight while every infrastructure dashboard reports the service healthy.
Tokenmaxxing read
This is the alerting spec teams keep meaning to write: page when estimated hourly spend triples against baseline, flag a workflow whose p95 token count drifts a fifth over budget, and chase truncation at 2% rather than waiting for 5%.
Source takeaway
Log model, workflow, prompt and completion tokens, latency, finish_reason and cost per call, using OTel GenAI semantic conventions instead of homegrown attribute names. Judge-grade only a 5% slice so evaluation does not double inference cost.
