Scaffolding, wordsmithing, and examples fix different failures.
Three fixes for a failing field. The right one depends on why it failed.
Scaffoldingmost of the time
Readers cited different evidence before reaching the field.
→Add an upstream field. Force the same anchor before the decision step.
Wordsmithsometimes
The question has a hidden ambiguity, or a term readers parse differently.
→Sharpen the wording. Name the threshold. Split the conjunction.
Examplefor edge cases
The question is clear, but a boundary case isn't.
→Show one paper's full decision path — quote, reasoning, answer. Not just the {input, answer} pair.
is_study_adequately_blinded→ fails (readers disagree). Which fix?
readers cited different evidence
scaffolding
add
participants_blinded, assessors_blinded, analysts_blinded upstreamsame evidence, different mapping
wordsmith
sharpen "adequately" to a specific threshold
same evidence, same rule, boundary case
example
show one paper's quote + reasoning + answer
Caveat on the example fix: if you lift an example from document d, keep d out of the testing loop — or you'll overstate generalizability.
When a field fails, first ask what kind of failure it is.
Different failures deserve different fixes.
@karlrohe