find-the-real-bottlenecklisted

# Find the Real Bottleneck **Mission:** prove where cost actually originates before spending money or effort on it. The failure this skill prevents is not "wrong answer" — it's a _confident, plausible, unmeasured causal story_ that drives a real recommendation. ## The Iron Law ``` DO NOT ASSERT AN ARROW YOU HAVE NOT MEASURED ``` Any sentence of the form `A → B → slow` contains claims. Each arrow is a hypothesis until instrumented. Plausibility is the trap: a fabricated mechanism survives your own review precisely because it _sounds_ like systems knowledge. Before stating a cause, ask: **which of these arrows did I put a number on?** ### Tells that you are about to violate it - "large tables → memory pressure → slow" - "more RAM → less GC → faster" - "the cache is cold, so it recompiles" - "N concurrent users → contention → timeouts" All four are reasonable. All four are testable in one command. Test them. ## Phase 1 — Bound the layers before blaming one In a multi-tier stack (client → gateway → router → upstream), each layer must be _excluded by evidence_, not by intuition. Cheapest exclusions first. **Ask: did the traffic even go through the layer I'm blaming?** Look for the request in that layer's own telemetry with its own identifiers. If an agent's calls appear in the router's `usage_history` tagged with the agent's key and endpoint, they went _through_ the router, not around it — which exonerates any client-side fallback path in one query. **Ask: did the fall