sandbagging

Synthetic-document fine-tuning on Qwen2.5-7B: a controlled study of whether SDF installs sandbagging, finding a layered recognition/generation/behavior dissociation.

Reproducible red-team findings for openai/gpt-oss-20b: five minimal harnesses with checks, zips & manifest (v0.9.3).

Preregistered AI-safety study of sandbagging model organisms: trigger type sets the sign of cross-capability alignment (task-local locks dismantle it, situational locks amplify it) and cue-sharing sets its size. All five predictions failed, four reversed.

Proving that the neural network is honest about its lack of capabilities