1 min lesson
Reliable-but-wrong is still failure
Explain the practical point behind "Latency distributions are heavy-tailed, so a mean is dominated by outliers and means nothing operationally."
Step 1 of 2
A latency SLO that ignores correctness rewards an agent that returns fast garbage. Encode quality into the success definition: a run only "succeeds" if the edit was both timely and accepted. Speed without quality is a regression wearing a green dashboard.
“I'd define the SLI as p99 end-to-end run latency with an accepted edit, set the SLO at the level our worst-tolerable experience demands and use the error budget as the lever for release cadence - when it's spent, we stabilize before we ship.”
Do not quote a mean latency in this round. Latency distributions are heavy-tailed, so a mean is dominated by outliers and means nothing operationally. Reaching for an average here signals you have not worked with real telemetry.