Of the pre-check's four answers, this page gives the second today: what it needs. Peak training memory at
your batch and the smallest card it fits, what one run costs, whether the graph will forward-pass at all, and the
layers, shapes and parameter count behind those numbers, computed from the structure before anything is trained.
Same card the bot posts on a pull request. What changed: /changes and
/compare. Layout and serving config: npx neurarch layout and
npx neurarch serve-check.