The authors introduce Ambient, a proactive egocentric assistance framework submitted to the EgoProactive track of the ECCV 2026 Wearable AI Challenge, securing first place in the large‑model division. By casting the assistance decision as a binary yes/no prediction and deriving it from renormalised token probabilities, they improve macro‑F1 by 0.249 and G‑mean by 0.30 over free‑form generation. To compensate for limited labeled data, they augment supervision with tool‑generated annotations.