Supplementary materials for the paper: Selective LLM Revision with Statistical Calibration for Sparse Retail Demand Forecasting

Supplementary materials for the paper "Selective LLM Revision with Statistical Calibration for Sparse Retail Demand Forecasting" submitted to the 16th International Conference on Computational Logistics (ICCL 2026), Hamburg. This document contains five appendices that were removed from the main paper to meet page limits: (A) The three intervention policy prompts in full (Conservative, Moderate, Aggressive), shown as instantiated for one representative product. (B) Reproducibility study details under deterministic decoding, including per-combination divergence statistics for a 200-product-store sample re-evaluated at temperature = 0. (C) Classical baseline comparisons against Croston SBA and regime-specialized Tweedie LightGBM, with per-regime breakdown of the LightGBM baseline. (D) Full per-flag-combination breakdown across all 14 flag combinations present in the test data, for the baseline and the three intervention policies under composition with the global shrinkage parameter. (E) Three example LLM reasoning traces drawn from the 1,665 Conservative-tier revisions: a deferring trace, a downward revision on a cold-start product, and a selective upward revision on a promotional window. Section, figure, and table references in this document refer to the main paper.