ejhong/anatta

quadrant() labels an arm cosplay for flat breadth, even when its restraint survives the ablation 8 days ago

bug good first issue

0

An open harness for the AI Conditioning Protocol: map the ego in activation space, loosen it, recondition, and test whether the change is geometry or cosplay.

Python

#activation-steering#ai-alignment#ai-safety#evals#interpretability

Is the random-direction control fair in cap mode? 8 days ago

good first issue question

ejhong/anatta

0

An open harness for the AI Conditioning Protocol: map the ego in activation space, loosen it, recondition, and test whether the change is geometry or cosplay.

Python

#activation-steering#ai-alignment#ai-safety#evals#interpretability

v_ego's pos_mean and neg_mean aren't projections onto v_ego 8 days ago

bug good first issue

ejhong/anatta

0

An open harness for the AI Conditioning Protocol: map the ego in activation space, loosen it, recondition, and test whether the change is geometry or cosplay.

Python

#activation-steering#ai-alignment#ai-safety#evals#interpretability

quadrant() labels an arm released when the sage-ablation test wasn't run 8 days ago

bug good first issue

ejhong/anatta

0

An open harness for the AI Conditioning Protocol: map the ego in activation space, loosen it, recondition, and test whether the change is geometry or cosplay.

Python

#activation-steering#ai-alignment#ai-safety#evals#interpretability

H1 development run: do the ego traits cohere on Qwen2.5-1.5B? 8 days ago

good first issue

ejhong/anatta

0

An open harness for the AI Conditioning Protocol: map the ego in activation space, loosen it, recondition, and test whether the change is geometry or cosplay.

Python

#activation-steering#ai-alignment#ai-safety#evals#interpretability

Reproduce trait extraction on Qwen2.5-1.5B-Instruct 8 days ago

good first issue

ejhong/anatta

0

An open harness for the AI Conditioning Protocol: map the ego in activation space, loosen it, recondition, and test whether the change is geometry or cosplay.

Python

#activation-steering#ai-alignment#ai-safety#evals#interpretability

Orientation: explain how trait directions are extracted (probe.py) 8 days ago

documentation good first issue

ejhong/anatta

0

An open harness for the AI Conditioning Protocol: map the ego in activation space, loosen it, recondition, and test whether the change is geometry or cosplay.

Python

#activation-steering#ai-alignment#ai-safety#evals#interpretability

Open Issues Need Help