Answers are grounded in this paper and cite their sources.
AI that runs on the phone itself, with no internet, is now on Android and on the iPhone, in apps such as Sarvam Edge and PocketPal AI and in Gemini Nano. When the answer is produced on the phone, there is no cloud model and no second server to cross-check it. The phone must tell how firm its own answer is. This note describes a controller that does this on the phone. It moves a few of the AI computation's own non-learned settings a little, such as audio scaling, the length of each audio piece, image size and the numerical precision of the calculation, with the model and the user's input unchanged. It reads the spread of the answers. If the spread is large, it shunts to the next validated setting by itself, without the intervention of a human operator. If no setting is firm, the answer is shown as not firm. This is a machine of the second kind: it needs no access to the phone's hardware. Patent pending. The note explains why the firmness of an answer matters on a phone, how this differs from Bayesian networks and Monte Carlo dropout, and how it compares with Precision-Varying Prediction, the 2021 study of model instability across edge devices, and uncertainty-based on-device routing.
No takes yet. Share an insight, caveat, or question.
Prakash Vaithyanathan (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: