Measuring What a Human Adds to a Machine Baseline: A Problem Statement

A working note, not a paper. Three studies in different domains asked whether a human's departure from a machine baseline carries information the baseline lacks: chess against a stronger engine, forecasting tournaments against language models, and noisy image classification against fine-tuned classifiers. They used four different operationalizations. This note states them as one quantity, the human's increment over the baseline within a combining family, distinguishes it from the increment relative to an ensemble reference and from the increment given both, and lists the open questions in the order they should be attacked. The results summarized here are reported in full in the three studies it cites.