2 min lesson
A one-pager that ships
Imagine this comes up at work. "Your take-home model says routing simple calls to a small model saves 15% of inference COGS. What single addition most increases a reviewer's trust in that claim?" Start with the practical move.
Step 1 of 2
Illustrative, built only on this section's assumptions.
Route simple tab and chat calls from the large tier to a small model first, then trim prompt context on the large-tier calls that remain. Routing alone saves about 15% of inference COGS on our assumptions of a 40% simple-call share and a small model at about 15% of the large tier's $/token, and the sensitivity table keeps the saving between roughly 12 and 18% when that share moves ±20%. The product risk is quality on calls we misroute, so we start with a conservative routing rule and a fallback to the large tier. We confirm it with $/request on routed traffic, paired with a quality signal such as acceptance rate on the same traffic, reviewed weekly for the first month.
Three failure modes sink most submissions. No stated assumptions, so no one can trust or challenge the numbers. A fence-sitting recommendation that lists five options and ranks none, which reads as analysis avoiding a decision. And numbers that don't reconcile to the known COGS total, which makes a reviewer distrust everything above them. Fix all three before polish.
Put the assumptions tab first, not buried last. Open your walkthrough with "here's what I assumed and where I'm least confident," then show the sensitivity table proving the recommendation holds even when those assumptions are wrong. Volunteering your own uncertainty reads as truth-seeking, and it disarms the obvious challenge before it's asked.