One common hallucination has a plain cause. The document does not say, and the reader answers anyway. It is not that the model misread the paper. There was nothing there to read, and no value it was allowed to write.
So give the reader an out. If the reader may write "not reported," it does not have to guess. This is a hallucination control that costs one line per field.
There are three kinds of absence, and they are not interchangeable. Collapsing them loses information you will want back at analysis time.
definitions
not reported — the question applies to this document and the document is silent on it.
not applicable — the question does not apply. A single-arm study has no randomization method to report.
ambiguous non-response — the document addresses the question but not responsively. It gestures at a power calculation without saying whether one was done. I have started calling this one out separately, because "not reported" hides it.
That a missing out increases fabrication has been measured. Rajpurkar, Jia and Liang built the canonical demonstration: extractive readers "can often locate the correct answer to a question in a context document, but they also tend to make unreliable guesses on questions for which the correct answer is not stated in the context."1 Slobodkin and colleagues ran the ablation the claim actually needs. Same model, same questions. One prompt mentions that a question might be unanswerable; the other does not. Detection of unanswerable questions improved by as much as 80 points.2
sources1Rajpurkar, Jia & Liang 2018, "Know What You Don't Know: Unanswerable Questions for SQuAD," ACL 2018 (doi:10.18653/v1/P18-2124). 2Slobodkin et al. 2023, "The Curious Case of Hallucinatory (Un)answerability," EMNLP 2023 (doi:10.18653/v1/2023.emnlp-main.220). Zero-shot; Flan-T5xxl, Flan-UL2, and OPT-IML.
The out is not free, which is a reason to name the absence types precisely. Slobodkin reports a cost on the other side too: about eight points of F1 on questions that were answerable. The reader abstains where it should have answered. The trade is strongly positive, but it is still a trade. A single vague escape hatch makes over-abstention easy. Three sharply defined values ask the reader to say which kind of absence it found.
Data Mint's codebook assistant should write these values in. Check that it did. Things get missed. When you read a new codebook back, this is one of the first things to look for.
The fallback goes in one tag. I put the allowed values in the Description and the fallback in **Default if missing:**. In a screening codebook that reads **Default if missing:** Not specified. In a trial-results codebook, **Default if missing:** Not reported. The reader sees the line and uses it.
see alsothe defaults are written in— What these values look like sitting in the codebook itself.a codebook, defined— Where the Default if missing tag sits inside a field.good instructions have good examples— Show the reader an example of an absence, not only of an answer.trust hasn't been deserved— Hallucination is the first objection people raise. This is one of the answers.