Transcribing handwritten deeds, wills and inventories from the Barbados archive for a Zindi competition. Fine-tuned TrOCR and Qwen2.5-VL models vote line by line through family-weighted minimum Bayes risk, and every gain was held to a noise floor measured for the leaderboard.
How it knows: No gain counted until it cleared a measured noise floor (about 1.5 thousandths on public ensemble deltas), and the ranked submission is trained from raw data in one notebook, with its saved weights reproducing the submitted file exactly.
Nicholas Smith - all work