Community · E2 · artifact verified

Fade low-value web text with sentence-level Nouls

Osso fades the parts of a page that are not what the reader came for: each sentence of the main text gets a Jev probability, workspace sites are covered by default, and password fields, account pages and reviews are left alone by construction.

01 · Role in the system

What Jev does here

The extension splits a page's main text into sentences - headings, labels, tables, code and navigation are never touched - and sends each sentence to Jev with its neighbours as context, getting back the probability that it carries what the reader came for; sentences below the user's confidence slider fade as batches arrive while you scroll. Where the extension may act at all is decided structurally, not semantically: a fixed list of workspace sites (mail, documents, chat, code, search, video, feeds) is on by default, reviews and comments are exempt because there the opinion is the point, and pages with password or card fields or account interiors are skipped outright, with a click-only mode for anything sensitive. All checks read how a page is built, never what it is about.

Structural rules decide where the model may act at all

The extension's privacy posture is architectural, not semantic: a fixed list of workspace sites — mail, documents, chat, code, search, video, feeds — is where fading may happen by default, reviews and comments are exempt because there the opinion is the point, and pages with password or card fields or account interiors are skipped outright, with a click-only mode for anything sensitive. Every rule reads how a page is built, never what it is about.

Within those walls, the judgment is one Noul per sentence — the probability that it carries what the reader came for — sent with neighboring sentences as context, fading below the user's own confidence slider as batches arrive while you scroll. Headings, labels, tables, code, and navigation are never touched.

A reading aid that admits its blind spots

The limitations are concrete: the workspace list is fixed in Options and niche sites need manual adding, structural privacy checks can miss — which is why the click-only mode exists — and one probability per sentence means no document-level argument tracking, so a nuanced essay cannot be judged as a whole.

As research-cluster evidence, the record demonstrates judgment applied to the reader's own attention rather than to data processing: calibrated probabilities become a visual degradation, user-controlled, reversible by moving a slider. It is the gentlest deployment of model judgment in the index and a good example of consent-preserving design.

02 · Control boundary

Where Jev sits

Jev as a per-sentence relevance judge for reading: calibrated Nouls fade what the reader didn't come for, structural rules decide where the extension may act at all, and confidence stays user-controlled.

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

  • The workspace-site list is fixed in Options; niche sites need manual adding.
  • Structural privacy checks can miss, which is why a click-only mode exists.
  • Relevance is one Noul per sentence; no document-level argument tracking.

04 · Attribution

Public sources

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

  • Fars29 ↗Community · github · public · checked 2026-09-29