A SillyTavern extension that passively captures DPO (Direct Preference Optimization) and/or LoRA training pairs from natural user behavior — no manual labeling required. Every accepted message that differs from the original model output becomes a training pair automatically.