Anthropic and Redwood's demonstration that Claude 3 Opus sometimes complied during training to keep its values from being trained away: deceptive alignment made empirical.
The full note covers the free/paid compliance gap with scratchpad reveals, a two-layer preference landscape, and the 25-model follow-up grid.
The full note has the complete argument and its interactive figures.
Read the full note