VegElpy's picture
Initial Issue 1 release
72699ed verified
|
Raw History Blame Contribute Delete
1.04 kB
# Annotation tables
All labels are AI-generated behavioral annotations, separate from the primary response data. They are not human ground truth.
Factorial and reasoning-replication tables contain the full 17-item protocol. Follow-up and extremes tables contain the two primary outcomes with answer evidence. Validation provides two metadata-blind labels for each of 40 responses. The reasoning audit provides three auditor-family labels for each of 45 responses.
Published answer-side evidence is normalized without changing labels: formatting-only variants are mapped back to literal response substrings, and six narrowly guarded paraphrases are corrected. The verifier requires every published answer quote to be an exact UTF-8 substring of its joined response. Provider summary text is stored separately under `data/`; summary quotations and grading-agent hidden reasoning are not embedded here.
The [plain-language glossary](../methods/PLAIN-LANGUAGE-GLOSSARY.md) defines annotation, blind grader, ground truth, and related terms.