10 users fine, 100 slow, 1,000 dead — the cliff was always there

AI-generated code optimizes for works, not for works at scale. Nothing is wrong until everything is wrong at once — and by then, it's launch day.

The failure is invisible right up until it isn't. The database query that returns in 50 milliseconds against 100 rows takes 30 seconds against 100,000. The API call fired on every page load is harmless with 10 users and a death spiral with 1,000. The code isn't broken — it was simply never designed to scale, and you tested it with a dataset small enough to hide every flaw.

That's the scaling cliff: performance doesn't degrade gracefully, it falls off an edge. You won't find the edge by watching the app run smoothly in dev. You find it by going looking — or your first real wave of users finds it for you, publicly, on the day it matters most.

Two checks from this doctrine

Real value, not a teaser — two of the 4 checks in full. The rest are in the one-pager.

The cliff exists whether or not you look for it. Find it in a load test, or your users find it for you.