Latency envelopes your PM can read

Product managers rarely dispute model accuracy once charts land in a deck; they argue about tail latency when a marketing push doubles traffic. We teach engineers to pair p95 graphs with plain-language envelopes: “Between 09:00 and 11:00 JST we stay under 180ms for batch size 32.”

Students rehearse outage narratives using historical traces from anonymized cohort data. The exercise is deliberately uncomfortable—some paths fail—because silence in class becomes noise in production.

We close with a tabletop review: finance, support, and engineering each annotate the same envelope document. The goal is shared vocabulary, not unanimous joy. A little productive friction is welcome.