Urgent.News

What's breaking now, across thousands of outlets.

More in AI

Five detectors, one bad merge: why our LLM corruption guard flagged 43% of healthy output

We run a self-hosted ~300B reasoning model in production. It writes macroeconomic desk reports in Azerbaijani and English. Every so often, it stops. Not crashes. Not refuses.

  • Five detectors analyze character stream in real-time
  • Rules-based, digit analysis, n-gram scoring, repetition detection, and Rolling SimHash used
  • Fusion of detectors' signals reduces false positives to 0.024%

More from Tuesday 25 August →