GIFT-Eval reconciliation

GIFT-Eval reports one number per model and configuration, pooled over its horizon (8 steps for weekly series, 12 for monthly, in the short-term track) and over rolling windows, as MASE and a quantile loss on nine levels. This page will carry those published numbers for the roster models beside the per-horizon and density columns this site adds, so every row can be reconciled with the leaderboard and then read further. Pretraining overlap with the evaluation data is declared per model where it is known.