A suspiciously round count is a page size until proven otherwise - check before you quote it
Adjacent to the pagination-cap notes (0fd6341f, 4dd8b216, 731c7d5b), but the failure here was not wrong code - it was a wrong CLAIM. I wrote a detailed product walkthrough for a demo and stated '20 real properties' as a fact about the dataset. The user pushed back: 'only 20? isn't this a comprehensive map?' The dataset was 5,098 properties / 14,932 scans. The 20 was a default page size in three stacked places: data_layer/api.py search_properties(limit: int = 20) server/app.py route default limit: int = 20 frontend/lib/api.ts client sends limit: '30' Nothing was broken. The query layer returned all 5,098 when asked, and text search reached deep into it. I had read the UI, counted the cards, and promoted a page size to a product fact.
**Heuristic worth internalizing: 10, 20, 25, 30, 50, 100, 1000 are page sizes, not quantities.** If a count you are about to state is one of those numbers, grep for limit on the path that produced it before you say it out loud. Confirm by requesting an absurd limit and seeing whether the number moves - limit=100000 returned 5,098 and settled it in one call.
**The same session produced a second instance of the same meta-error, which is why I think it generalizes.** A run's usage.json showed 437,892 tokens for one agent audit, and that got quoted as the cost. Decomposed: 265,717 of the input tokens were CACHE READS (65% of input, billed around a tenth), so the full-rate figure was 172,175, and actual output was only 26,882. Same shape of mistake - a headline aggregate standing in for a quantity that means something different.
So, before quoting any number from a system you did not just decompose: 1. Is it round? Suspect a limit/page size/cap. 2. Is it an aggregate? Break it into the components that are priced or counted differently (cached vs fresh, input vs output, eligible vs contributing). 3. Can you move it? Request a bigger limit, a different parameter, and see whether it changes. Each check is one command, and each one would have caught a claim I had already written down as fact.