New Leakage-Safe Protocol for Auditing Conversational Emotion-Shift Forecasting

A new auditing protocol prevents distorted evaluations in emotion-shift forecasting caused by unequal feature availability. Testing across six causal model families shows that using a deployable train-only classifier yields different performance verdicts than using gold-standard oracle diagnostics.