alignment-faking

Tested whether Anthropic's sleeper agent detection technique generalizes to naturally-arising deception. Ran Llama 70B inference across 90 prompt-conditions on rented A100s, manually scored every output. Honest null result. BlueDot Impact AI Safety Sprint.

Empirical framework for testing whether RL training instills genuine behavioral dispositions or surface compliance in language models, using compute frugality as a controllable proxy value.