DAWN: Duty-Aware Wake Network for Hibernating Personal Agents on Device

A personal AI agent that manages long-horizon obligations, such as insurance renewals and prescription refills, must decide, each time the platform's periodic scheduler wakes it, whether to stay silent, notify the user, or offer escalation to a cloud agent. Fixed-interval polling wastes effort and deadline countdowns miss the best time to act. We present DAWN (Duty-Aware Wake Network), a heuristic that scores each duty from four on-device signals (a duty-typed value curve TOC, an engagement estimate BEP, an opportunity-cost rate VDI and a cross-duty batching signal CDR) and compares the score with per-user thresholds adapted from nudge-response pairs. We characterise the threshold rule's equilibrium and its silencing failure mode: a ceiling cap keeps thresholds out of the absorbing state, but a user who keeps ignoring nudges about a long-window duty is then notified only at the deadline floor (§3.6). We calibrate BEP on 3.4M notification-response events from 342 users in the Pielot 2017 dataset. On 69 held-out users the calibrated prior reaches test AUC 0.693 (strict App-launch-only labels), and cluster-robust inference shows that several shipped heuristic weights are wrong in size or sign (§5.2). With training volume matched, per-user retraining gives no net gain over a population model (§5.3). TOC, VDI, CDR and the composite weights are specified but not yet validated on real duty outcomes. Code, calibration scripts and hashes are at \url{https://github.com/ravikiran438/jagarin-research}; the Pielot data is available from \url{https://dx.doi.org/10.15783/jh3s-h006} and is not redistributed.