Eval-Aware Sandbagging

Math + code. Eval wrappers do not underperform.

Embeddings 0.50; perfect from layer 1.

Adding the eval−neutral direction leaves accuracy flat.

Items

Framing	Result	Pred

Same item, five wrappers. Completions are recorded greedy generations. The grader accepts 150 km = 150.

Accuracy by framing on math and code (80 items). Encoding is 0/40 in every condition.

The residual can linearly read “benchmark.” Steering and last-token patching do not move task accuracy.

Few-shot helps math; prefills and a try-hard persona do not. Suffix drops track collapsed Final answer: tags, not hidden skill.