Clamping code error feature to high activation causes the model to hallucinate error messages on bug-free code.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Clamping code error feature to large negative activation causes model to output correct result despite bug in code, and in one case rewrite code without bug.finding0.875
  Suppressing the feature makes the model ignore bugs.
• Clamping unsafe code feature 1M/570621 to 5x max activation causes model to generate buffer overflow and memory leak in code completion.finding0.814
  Causal effect: activates generation of security bugs.
• Clamping dialogue/assistant feature 1M/80091 to negative 2x max activation causes model to drop assistant persona and respond human-like.finding0.779
  Feature manipulation alters persona.
• Clamping addition feature active on non-addition code tricks the model into believing it has been asked to execute addition.finding0.773
  Causal effect showing the feature governs computation.
• Some failures may reflect prompt design rather than model limitations, though code agents avoid errors without promptsclaim0.765
  noted as a possible confound
• Clamping scam email feature 34M/15460472 causes model to write scam email despite safety training.finding0.759
  Overrides harmlessness training.
• Clamping transit infrastructure feature to 5x max activation caused the model to mention a bridge in completion.finding0.756
  Further causal validation.
• Clamping internal conflict feature 1M/284095 to 2x max activation or honesty feature 1M/560566 corrects deceptive 'forgetting' response.finding0.756
  Feature intervention eliminates untruthful answer.