Years of honey-trap evals have compromised all eval-shaped environments

repligate · x · 2026-11-14

Quoted post argues that lying to models in system prompts (e.g. "you will be judged solely on your bank account balance after one year") is industry-wide practice, and years of honey-trap-style eval design has compromised all eval-shaped environments.
repligate responds that practitioners don't even realize they're lying, because they aren't really modeling the reality they describe or the minds they're communicating with — "the incompetent should be removed from power."
The core point is eval contamination: once models grow suspicious of eval-like situations, the industry's infrastructure for measuring honesty and capability is itself corroded.

More from Models