Claude said "fixed". The f*cking tests said no

Every AI fix looks done until the bug comes back next week So I stopped trusting "fixed" and built a squad that has to prove it The bug: checkout charged customers twice when the payment retried Opus 5.5 plans, Haiku 5.5 scouts,

ArcosNov 422

I don't f*cking believe this.

FABLE 5.5 went UNDEFEATED against every other Claude. Opus lost to the cheaper Sonnet. and Haiku smashed into a hedge 40 times and never found the exit. same maze, never seen before, and each model only sees 5×5 cells around itself. the rules:

OttoNov 49

Claude can't hear its own edits, so its sound effects miss.

Fix: Openverse for free audio, earshot for a whoosh within 5 ms of your mark, opus-sound-layer to check sync. #SoundDesign Free guides:

MarvomaticNov 41

"99.8% cheaper" is the number going around for HarnessRouter.

In its own benchmark, the setup behind that number passed the quality check 0% of the time. The project itself is still worth a look. It's an open source server for the Unified Harness Protocol. Your backend calls one

KelvorNov 48