Filter your X timeline with five typed signals

01 · Role in the system

What Jev does here

Each eligible post on the timeline, together with the reader's interest topics, is sent to Jev as one batched request carrying a rubric of five noul questions: relevance to the reader's interests, substance, practical value, promotion, and engagement bait. A fixed preamble tells Jev the post and topics are untrusted data rather than instructions and not to infer the content of links. The extension then combines the returned signals with reader-owned weights and thresholds entirely on the device, maps the composite to a reversible presentation - highlight, label, dim, collapse, or hide - and caches each verdict for a day keyed by a hash of the request, with concurrency slots and a local daily limit so the feed never blocks on the model.

A rubric of five judgments, not a blocklist

The rubric judges what a post is, not which words it contains. Its five noul questions each name an outcome: relevance to at least one of the reader's stated topics, at least one substantial concrete detail such as an explanation, measurement, or identifiable source, a practical step the reader could apply, whether the primary purpose is to sell or solicit, and whether engagement-seeking tactics dominate. The criteria are written to resist over-blocking: a useful release note or technical explanation is not automatically promotion, and disagreement, emotion, humor, or criticism alone are not engagement bait.

The reader owns the tuning, the feed stays reversible

Every signal feeds a weight the reader sets, alongside a threshold, a highlight threshold, a minimum margin, and a behavior for posts that fall below the line. Three profiles ship built in - Builder, Curiosity, and Deep focus - and each is just a saved set of those numbers, so tuning is auditable and reversible in the same way the feed is. Presentation changes never delete anything: collapsed posts keep a Peek interaction, and hiding is a setting the reader can switch off, not a judgment the extension locks in.

A privacy boundary written into the code

The extension has one connection path: the page to a service worker to the TypeSafe Jev API, BYOK, with the key session-only by default and persistence an explicit opt-in. Host permission for api.typesafe.ai is requested only at connect time and is revoked when the key is removed, and the README documents the boundary in plain words - this is extension software, not offline AI, and TypeSafe processes the text remotely under its own terms. There is no intermediary backend, no registration, no analytics, and the rubric's preamble is the standing defense against a timeline that tries to talk its way out of being judged.

02 · Control boundary

Where Jev sits

Visible post text plus reader interests become one batched noul rubric per post; numeric signals return, and deterministic local code owns the weighting, the thresholds, and every presentation change.

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

• Jev sees only the post text: the preamble forbids inferring the content of links, images, and videos, so a bare-link post is judged by its words alone.
• The signals are judgments of primary purpose, not fact checks: promotion and bait scores say what a post is for, not whether it is true.
• Filtering sends eligible text and interests to TypeSafe under its own terms; the boundary is BYOK with no intermediary, but it is remote inference, not local.
• On request failure or the local daily limit the feed is left untouched, so a degraded API degrades filtering, not the timeline.

04 · Attribution

Public sources

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

• MithrilMan ↗Community · github · public · checked 2026-09-20
• MithrilMan ↗Community · demo · public · checked 2026-09-20