CASOONresearch project · in development
A website is a surface. Relief makes its structure tangible.
Relief is a research browser built on Chromium. It reads the accessibility tree inside the browser, turns it into a semantic model of the page and lets people operate the page by intent — alongside VoiceOver or NVDA, not instead of them.
- p95 from accessibility packet to updated page model
- ≤ 1.1 ms
- small patches outside Relief’s own directory in the Chromium fork
- 9
- recorded pages with correct page type and primary action
- 13 / 13
- VoiceOver output with Relief running, compared step by step
- unchanged
What Relief is built to do
Understand the page, not just read it
Page type, functional groups, the primary action and what changed — derived from Chromium’s own accessibility tree, including cross-site iframes such as consent dialogs.
Operate by intent, safely
“What can I do here?”, “add size 43 to the cart”. Every action runs through a validated plan in Rust; risky actions need an explicit confirmation. A model may suggest, never act.
Say where a statement comes from
Every answer carries its certainty: known from the page, inferred, or uncertain. Guesses are never presented as facts.
Test accessibility the way people use it
The same core runs as a test mode: form journeys, focus after errors, live announcements and the real VoiceOver output, checked in the real browser.
Where the project stands
Phase 0 — feasibility inside a real Chromium build — is complete. The user-facing surfaces of phase 1 now run in the fork.
- Done: Chromium fork builds on Apple silicon; accessibility tree → Rust → page model → action → result, measured on real sites. VoiceOver keeps working unchanged next to Relief.
- Done: actions through Chromium’s accessibility actions, an inspector with findings, a command bar, keyboard hints, a form assistant, consent handling and a semantic view.
- Done: speech — a talk button with live microphone, recognition on the device, spoken answers; “cancel” interrupts at once.
- Done: ability profiles — output, input and presentation adapted to abilities, globally or per site, stored locally only.
- Done for testing: form assertions, headless runs with JUnit reports, Playwright access to the page model.
- Next: real screen-reader output in tests, an accessibility baseline for Relief’s own interface checked with VoiceOver, the runtime in a sandboxed utility process.
Frequently asked questions
Does Relief replace a screen reader like VoiceOver or NVDA?
No. Relief runs alongside VoiceOver and NVDA rather than instead of them, and the Chromium fork keeps VoiceOver output unchanged while Relief is active.
What parts of Relief are actually usable today?
Phase 0 — feasibility inside a real Chromium build — is complete. The phase 1 surfaces run in the fork: the Relief panel, command bar, semantic inspector, semantic view, speech, bookmarks, the form assistant and consent handling.
How does Relief handle cookie and consent dialogs?
It detects consent dialogs, including those in cross-site iframes, and can decline them on request — it never accepts a consent dialog on a person’s behalf.
Can a model take actions on a page by itself?
No. Every action runs through a validated plan in Rust before it reaches the page, and a model may only suggest an action, never execute one directly; risky actions need an explicit confirmation.
Can Relief’s behavior be tested automatically?
Yes. The same core runs as a test mode with form assertions, headless runs that produce a JUnit report, and Playwright access through the CDP domain Relief.*.