Preregistration: Do LLM Pipeline Evaluations Notice Broken Preprocessing? A Mutation Analysis of Extraction Scorers
Preregistration (version 1.0) of a mutation study of LLM extraction pipelines on CUAD v1. Twenty-two mutation operators break stages outside the model (PDF reader, cleaner, chunker, prompt, parser, merge), and the study measures how often seven common scorers (exact match, SQuAD token F1, CUAD Jaccard, token-set, fuzzy, embedding similarity, LLM judge) fail to notice that the output changed, and whether three label-free stage checks catch what they miss. Posted before any study-set document was drawn or run. The zip holds the frozen code, the stratum file and the disclosed development dry run (5 contracts). sha256 of scorer-mutation-study_code_v1.0.zip: 5a5d251e6af2a82b458d0edc7dd2f424da46cd84f9de30779e2f6ddaf9960d4e
Authors
- Arshdeep Ghotra
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24409116
- Primary Topic
- Topic Modeling
- Type
- preprint