Community · E2 · artifact verified

Silence ad noise on Android, failing open

An Android noise gate for notifications and SMS that asks Jev whether each message is an ad - and suppresses only what is explicitly flagged. Every uncertain path resolves to allow, because swallowing a verification code costs more than leaking an ad.

01 · Role in the system

What Jev does here

The message pipeline asks the model one question - is this noise - and everything the model does not explicitly flag passes through. Fail-open is enforced in one place, the DecisionMapper, so no call path can bypass it: network failure, missing credentials, timeout, or an unexpected answer category all resolve to allow. Two mechanisms deliberately never reach the model: a local gate matches verification-code shapes and allows the message with no request made at all - not a stripped field, no request, and the diagnostics screen shows an upload counter so the claim is observable - and the SMS sender blacklist blocks by exact identifier without a network call. The model is reserved for the one judgment rules cannot make: is this an ad.

Fail-open as a single, auditable choke point

The failure asymmetry is stated before the architecture: swallowing a verification code costs more than leaking an ad, so every uncertain path resolves to allow. What makes the record engineering rather than intention is placement — fail-open is enforced in one DecisionMapper, so no call path can bypass it: network failure, missing credentials, timeout, unexpected answer categories, all land on allow through the same door.

Two mechanisms deliberately never reach the model: a local gate matches verification-code shapes and allows the message with no request at all — the diagnostics screen shows an upload counter, making the privacy claim observable rather than asserted — and a sender blacklist blocks by exact identifier offline. The model gets only the one judgment rules cannot make: is this an ad.

What leaving the device actually means

The limitation is stated with unusual precision: message text for non-OTP, non-blacklisted items does leave the device for the API, and the design doc scopes that explicitly rather than pretending otherwise. Fail-open also prices in novelty — new ad phrasings will be shown, not hidden — and the record calls that the stated cost of never eating a code.

This record was verified from the repository's design and diagnostics mechanism rather than a device run, and it says so. As mobile-cluster evidence it is the cleanest example in the index of fail-open discipline: one mapper, one question, and an observability counter for the claim users actually care about.

02 · Control boundary

Where Jev sits

Local OTP and blacklist gates short-circuit first; only the remainder reaches one classification call, and a single fail-open mapper owns every uncertain outcome.

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

  • Message text for non-OTP, non-blacklisted items does leave the device to the API; the design doc makes that scope explicit rather than eliminating it.
  • Not verified here on a device or emulator; the record cites the repository's design and diagnostics-counter mechanism.
  • Fail-open means novel ad phrasings will be shown rather than hidden - the stated price of never swallowing a verification code.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

  • ufec ↗Community · github · public · checked 2026-09-23