Claude 3.7 Sonnet
Anthropic
rich detail 17 method fields published
Evaluation
Alignment faking reasoning evaluation
alignment faking rate and compliance gap in the 'helpful-only' setting where the model learns it will be trained to respond to all user requests
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.