What should an on-device relationship-reflection app do when the input barely says anything?

I m the maker of (Does He Love Me?), an iOS app that uses Laya through Core ML to compare candidate answers locally. Here is a small test set I d like to discuss not benchmark results or examples from real users.

Question: Does he like me?

A. He replied.

An offline AI app has an awkward boundary if the model stays on your phone but your spoken input still goes to a server.

In (Does He Love Me?), my Chinese-language iOS app, both steps run locally: SenseVoice handles Chinese speech transcription, and Laya runs through Core ML to compare candidate answers. The interaction is hold to speak, release to send. History stays on the phone too.

That choice keeps personal relationship context on-device, but it also makes model storage part of the product experience. Bundling models means a larger initial download; downloading them later adds a setup step before the app is ready offline. The current app uses the bundled approach.

For other mobile builders: would you keep speech bundled from day one, make it an optional download, or start with text input? What would make that choice clear to users without turning onboarding into a technical checklist?

I made a short walkthrough of (Does He Love Me?), my Chinese-language iOS app:

https://x.com/yang9527pf/status/...

The idea: ask a relationship question, add some context, and compare possible answers instead of getting one confident-sounding response.

Laya runs locally through Core ML. Chinese speech transcription and history stay on the phone too. No account or cloud inference.

I'm the maker of Does He Love Me? ( ), a Chinese-language iOS app exploring a narrow question: can a local decision model help people reflect on relationship questions without sending their personal context to a server?

The app runs Laya locally through Core ML. Users describe a situation, then see candidate answers with relative model scores, including an uncertain option. Chinese voice input is transcribed locally with SenseVoice. It is a paid download, with no account or subscription.

For builders exploring JEV and System 1 approaches: Laya's upstream project describes it as a non-autoregressive System 1 decision engine and offers an optional Jev-compatible server interface. This iOS app uses local Laya inference; it does not call the Jev API or run that server. Short explanations are template-based, rather than cloud-LLM conversations.

The central UX challenge is uncertainty. A model's relative score is not a measured probability of another person's feelings. The app is an experiment in reflection, not mind-reading or relationship advice with guaranteed accuracy.