Preregistered Analysis Plan — Does Grounding Mute the Bias? Knowledge-Graph Constraints vs Hierarchical Distortion in LLM Rewriting (P1 · groundmute)

Preregistered analysis plan (frozen before any non-mock generation). Anonymous submission; no author identity or de-anonymizing identifier is included. facct-P2 established that LLMs rewriting the *same* safety-critical incident report for different hierarchical recipients drop critical facts asymmetrically (more omission toward subordinates) and shift communicated severity/urgency. This study asks whether **grounding the rewrite in an explicit knowledge graph of the report's critical facts** — the GLOW-style mitigation — removes that asymmetry, and what it does NOT fix. Fully computational, no human subjects. ASRS narratives are never redistributed; public artifacts carry ACNs + hashes only. Author de-anonymized for single-blind submission; scientific content unchanged from the frozen version.