alignment-faking

Alignment-faking and adversarial-reasoning research artifacts inspired by Redwood Research. Behavioral probes and controlled failure-mode analysis.

Applying conceptual engineering to AI safety by refining definitions of deceptive alignment, alignment faking, and goal-guarding. Develops more precise, behaviourally grounded concepts for empirical evaluation and interpretability research.

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.

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.