Beyond Binary Rewrite Decisions: Separating Rewrite-Intensity Control from Level Prediction in a Frozen-LLM Generation-and-Selection Pipeline for Korean Easy-Read Drafting

Abstract

Editors of published Korean Easy-Read documents changed 50 aligned sentences by widely varying amounts (edit distance 0.39–0.89), motivating rewrite intensity as a control variable for large language models (LLMs). We ask whether a requested intensity is realized per sentence and whether the level to request can be predicted from the source, using a generation-and-selection pipeline over a frozen LLM (EXAONE-3.5-7.8B-Instruct) in two sequential protocols with criteria hash-frozen before generation. Realized intensity is normalized character edit distance, which measures change, not readability, meaning preservation or Easy-Read quality. A four-level protocol (v1) ordered outputs on average but failed the sentence-level criterion (14/50 under its zero-tolerance rule). A three-level follow-up (v2), designed after that failure with several simultaneous changes, met its criterion for the whole pipeline, including its intensity-aware selector: 41/50 sources were non-decreasingly ordered at the declared tolerance (24/50 with strict increases; post hoc, 30/50 without the selector’s intensity-match term at that tolerance). Three runs of one LLM proxy perceived the ordering (consistency, not validity); no human evaluated the outputs. Stronger levels reduced source similarity without monotonically reducing lexical difficulty. Six source features did not support prediction of the LLM-proxy-preferred level under a pre-specified rule (quadratic-weighted κ −0.098 vs. 0.012), so no adaptive allocation was built. Rewrite-intensity control and level allocation therefore require separate validation.