Disagreements should inspire refinements to the codebook.

Ambiguity in the codebook becomes disagreement in the data. The disagreement is what points you back to the ambiguity.

Tempting but usually wrong: "the reader was bad; swap models." Almost always, the fix is in the codebook.

You can build your trust in your codebook by diagnosing the ambiguities that created the disagreements, iteratively refining the codebook, and then seeing the readers align.