Community · E2 · artifact verified
Watch your screen time with local and hosted judges
Qualm is a screen-time app for Apple silicon Macs, local-first, built on two judges: Kev running on your machine and TypeSafe's hosted Jev - typed choices decide what counts as a distracting session without your activity leaving the machine unless you opt in.
01 · Role in the system
What Jev does here
Screen-time classification is usually a pile of regexes. Here each session becomes a typed choice - which app category, focused or distracted - answered by a decision model: Kev locally for the default path, Jev hosted for judgments you route to the cloud, both returning calibrated probabilities rather than tag strings. Local-first means the Apple-silicon path works with no network and no per-call cost; the hosted judge is opt-in per the user's choice, and the trial layer records what was decided and by which judge, so your screen-time report is an audit of decisions, not a black box.
Two judges, and the private one is the default
Screen-time classification is usually regexes over app names. Qualm turns each session into a typed choice — which category, focused or distracted — and answers it with whichever judge the user chose: Kev running locally on Apple silicon for the default path, TypeSafe's hosted Jev for judgments explicitly routed to the cloud. Local-first here is not a slogan; the local path works with no network and no per-call cost, and the hosted path is opt-in.
The trial layer records what was decided and by which judge, so the screen-time report doubles as an audit of decisions. That is a privacy architecture most hosted-classifier products never surface: you can see which judgments left the machine, and what came back.
The unpublished number that matters
The honest gap is named in the limitations: agreement between the local and hosted judges is not published. A dual-judge architecture invites the question — do Kev and Jev actually classify the same session the same way? — and the record does not answer it yet.
Scope limits are otherwise clear: Apple silicon Macs only, the local path needs a model download, and category criteria are the author's, not yet user-tunable. As the home cluster's only dual-judge record, it is the natural starting point for anyone comparing local and hosted decision models on identical inputs.
02 · Control boundary
Where Jev sits
Jev as the hosted half of a dual-judge screen-time classifier: sessions become typed choices answered locally by Kev or in the cloud by Jev, every decision recorded with its judge and probability, defaulting to the path that never leaves the machine.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Apple silicon Macs only; the local Kev path needs the model downloaded.
- 4 stars; category criteria are the author's and not user-tunable yet.
- Dual-judge agreement between Kev and Jev is not published.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- RoderickQiu ↗Community · github · public · checked 2026-10-03