On-device continual learning gets an artifact: 230M model, rank-4 LoRA, one frozen teacher

The economical move here is the absence of a second model. The deployed adapter acts; the same frozen base weights then re-read the finished interaction with the adapter switched off and the observed outcome appended to the original context ^[8][7]. Teacher and student are one network, so the only advantage the teacher holds is the outcome sitting in its input ^[1]. On a handset that is the difference between one resident set of weights and two: reviewing an interaction costs an extra forward pass, not a second model in memory ^[2].

That also explains why most of the design effort goes into what not to learn from. Supervised fine-tuning wants a correct action the phone never observes, and a reward is no better defined, since an opened notification is not proof of good timing and an ignored one is not proof of unimportance ^[12]. The obvious repair, mapping a dismissal onto "archive", overstates what a dismissal means, which could be irrelevance, bad timing, a preview that already did the job, or an interruption ^[13]. So outcomes arrive as reliability-conditioned soft targets: strong support for one action when the outcome is reliable, redistribution across the still-plausible actions when it is ambiguous, and no update at all when it reveals nothing ^[9]. The pseudocode puts a causal-support check in front of the buffer as well ^[10]. Two gates stand between a user gesture and a gradient step ^[3].

The limit the design cannot engineer away is the branch not taken. Archive a notification and the result of showing it is unobservable, to the reviewing pass as much as to the acting one ^[5]. Hindsight is richer than a label and still one-armed. The prototype's answer is exploration while the model is uncertain, tapered as confidence grows, on the argument that a greedy policy locks into an early behaviour and then collects only the evidence that confirms it ^[11]. Look at what the taper does to the data supply: the mechanism that produces off-policy evidence is scheduled to wind down precisely when the policy has become confident, which is also when a change in the user's routine would be hardest to notice. The bounded replay window and the category-balanced sampling are the counterweight ^[14], and whether they are enough is an empirical question.

Which is where the excerpt stops. There is no routing comparison with the adapter learning against the adapter frozen, no per-step latency, no battery cost ^[4]. For a training loop that runs inside a thermal budget, those are the numbers that decide whether it ships or stays a demo.

What the artifact does settle is narrower and still worth having. A 230M model with a rank-4 adapter, ONNX Runtime Training and a bounded buffer is enough to keep updating after deployment, with provisioning the only remote step described ^[2][3]. The frozen-at-ship default that shipped small models inherit ^[4] now needs an argument other than the hardware.