State that the base model contributed nothing measurable

Replace the ModernBERT calibrator with one fitted on VERA

Correct the pair set, the training counts, the chance rates and the calibrator

Add measured LLM and Laya baselines, and the Reddit transfer result

Correct the prior-work table and add the significance-stratified accuracy

Add library_name, widget, model-index and DOI citation

Tighten the card: one fact per sentence, active voice

VERA: Bradley-Terry ranker trained on 62,695 measured A/B arms

initial commit