Equivalence of structured and original prompt formats

Establish whether the structured answer-field rendering is equivalent to the original candidate-aggregation prompt for purposes of measuring accuracy and causal answer-field effects.

Background

The answer-field intervention changes the candidate presentation by adding an explicit reported-final-answer field and manipulating that field while retaining identical reasoning prefixes. The intervention finds that substituted answer values steer outputs, but masking does not measurably recover accuracy.

Because the structured rendering differs from the original prompt, the authors conduct a bridge comparison. Its confidence interval is not contained within the pre-fixed equivalence margin, so the paper cannot transfer the structured intervention's causal conclusions directly to the original-format experiment.