Language Twin: A Shared-State Architecture for Terminology-Consistent Document Translation with Human-Edit Propagation: A Pilot Study

Large language model (LLM)-based document translation systems typically treat each segment independently, discarding terminology decisions, human corrections, and discourse cues after each generation step. This stateless approach causes terminology inconsistency across segments, failure to propagate approved post-edits downstream, and redundant prompt-token consumption. Existing solutions—document-level MT, retrieval-augmented generation, and computer-assisted translation (CAT) tools as a general category—address individual aspects but lack a unified, state-aware architecture with provenance, update rules, and rollback semantics. We propose Language Twin, a shared-state architecture that organizes translation projects into seven versioned layers (L0–L6), supporting selective context loading, scoped human-edit propagation, and reversible updates. A pilot study translated three curated English-to-Korean document bundles (17 segments) using GPT-4o with a temperature of 0.3. The Language Twin condition (P1) achieved numerically higher preferred-term accuracy than the strongest baseline (17/21 vs. 14/21; not statistically significant at this sample size) and showed no repeated downstream errors in the monitored set (0/5 vs. 5/5 against the propagation-disabled ablation; Fisher’s exact test: p = 0.008), while reducing prompt tokens by 39.2% relative to full-context loading (A4). In blinded human evaluation (quadratic-weighted κ = 0.71–0.78), P1 achieved the highest terminology rating (4.38/5 vs. 3.97/5) and lowest post-editing time (16.9 s vs. 19.1 s per segment). These pilot-scale results indicate that governed shared state can improve terminology consistency and editing efficiency.