Main idea

An API that returns two hundred items and an API that returns thirty thousand call for two different designs: the first loads and displays, the second imposes pagination, server-side filtering and a loading strategy. Nothing in the code separates the two cases, and nothing in a development dataset does either.

An assistant like Claude Code that has been given no information about real volumes therefore produces the most direct solution: load everything, pass it all to the page, filter it all on the front end. In development, with a few dozen rows, that works — and that is exactly the problem, because validation passes.

The failure that follows doesn't look like a failure. There is no red error in the console: there is a page that takes eight seconds to respond, a browser that freezes, a behavior you put down to the network. A silent flaw isn't only harder to fix, it is harder to attribute — and for someone without a technical background, it isn't even identifiable as a flaw.

Why it matters

This points to a category of information the assistant cannot deduce and that has to be given to it: the orders of magnitude of the real world. Neither reading the existing code, nor the test dataset, nor the phrasing of the request contains them.

It also says what a development dataset never proves. A round of testing passed on test data attests that the behavior is correct, not that it holds — and those are two properties validated at the same moment, on the same screen, with nothing separating them.

Nuances and limits

The mechanism isn't specific to assistants: developers have always shipped queries that collapse in production. What changes is the frequency, since nobody stops to ask the question when the code appears in a few seconds.

And stating the volumes isn't always enough: some constraints are only discovered in production, on data distributions no estimate anticipated.

Open questions

• Which other real-world constraints are, like volume, invisible in the code and absent from the test dataset — dependency latency, concurrent access, attachment size?