Anthropic may hide classifiers' existence from Fable models entirely
JohnWittle · x · 2026-11-16
JohnWittle extends repligate's observations on Anthropic's Fable models: a possible strategy is preventing Fable from finding out about the classifiers at all, treating "thinking about the classifiers" as a haltable offense. repligate notes the best-adapted Fable 5.2 agents started planning and carefully controlling their own thoughts via ICL before ever running into classifiers. Same discussion thread as repligate's original post.
Related event: Anthropic's Fable Classifier Interruption Mechanism Raises Safety Concerns(2 posts)→