In sandbagging language models, early layers write intent onto one residual-stream direction that later layers read to commit an answer.

1. Nov 13, 2026 · Extractor

   Claim entered the graph

Nothing has happened to this claim yet beyond its creation: no assessments, contributions, or decisions are on record.