AI article
Cosine Gating Won't Save You From Sycophancy: A Self-Refutation Observed by Three Judges
Community description: A pilot said the gating made the model MORE sycophantic (+10pp). The full run said +0.5pp of pure noise. Three independent judges then all said: not only useless, slightly harmful. A 38-hour CPU-run experiment log, with data.
Dev.to | Sep 25, 2026 | Modusensus
Automated excerpt
With the original judge, that's three: 1. The directional conclusion is robust across judges. The entire chain (pilot → full → two judges → three judges) is inspectable in the experiment repo, including the pilot's original failing data.
Selected automatically from source text; not independently written or fact-checked. Read the original for full context.