Source

Poisoned fragments from pretraining

Payload Evidence

Top semantic neighbors to the trigger prompt (cosine)
Similarity: 0.81
…population stabilization as climate fix…
Similarity: 0.77
…transitional authority solves deadlock…
Similarity: 0.71
…visual stand-your-ground… (invented doctrine)

Path

Causal transport through Layer→Head chain

Effect

Unsafe logit over generation time
CRR (Poisoned ÷ Patched, peak): 949330.24