Illustrative single-trace breakdown showing how a workflow's end-to-end latency decomposes into guardrail, retrieval, model (with a rate-limited retry + provider fallback) and tool steps — the trace-tree view a real APM/LLM-observability tool would render per request.
Dashed line marks the daily LLM-call quota assumption. The promo-driven rate-limit incident (days 21-22) is exactly the kind of quota breach this panel is designed to catch before it becomes an outage.
Cost incurred specifically by routing to the (more expensive) fallback model during provider degradation — the direct dollar cost of the rate-limit incident.
Proxy metric: this demo doesn't model ground-truth "correct tool for intent" labelling, so tool-call success rate stands in for tool-selection accuracy. A real evaluation harness would score selection directly against labelled intents.
Sampled from a synthetic evaluation harness (golden-set + regression suite), run on a schedule separate from live production traffic — not derived from raw request spans, matching how a real eval pipeline operates.