PointCheck is a new tool that enhances web accessibility for users with disabilities.
Concept 1: What is Web Accessibility?
The Big Idea: Web accessibility means designing websites so that people with disabilities can use them effectively.
Key Details:
- People with low vision use browser zoom or high-contrast modes
- People with motor impairments may navigate entirely by keyboard (using the Tab key to move between elements)
- When websites aren't built with these users in mind, their assistive tools break down
Why It Matters: For many people, an inaccessible website isn't just frustrating — it can mean being cut off from transportation, healthcare, education, or social connection.
Concept 2: The Limitation of Traditional Accessibility Testing
The Big Idea: Most accessibility software only checks the underlying code, not what the user actually sees on screen.
The Problem:
- These tools compare HTML code against standard accessibility guidelines
- But compliant code can still produce an unusable page
- There is a gap between what the code says and what the user experiences
Simple Analogy:
A recipe can follow all the rules of cooking, but the dish might still taste terrible. Checking the recipe isn't the same as tasting the food.
Concept 3: Focus Indicators — A Critical Accessibility Feature
The Big Idea: A focus indicator is the visual highlight (usually a colored ring or outline) that appears around a button or link when a keyboard user tabs to it.
Why It's Essential:
- It tells keyboard-only users where they are on a page
- It's the equivalent of seeing your mouse cursor move
- Without it, a user can't tell if they're on "Buy Now" or "Cancel"
The Hidden Problem:
- A focus indicator can pass every automated code check while being completely invisible against the background color it's rendered on
- This is exactly the kind of flaw traditional tools miss
Concept 4: Browser Automation with Playwright
The Big Idea: Playwright is an open-source tool that can control a web browser automatically — like a robot user.
How PointCheck Uses It:
- Loads a target webpage in a background browser window
- Tabs through the page exactly like a keyboard user would
- Takes a screenshot at every step
Pre-Scan Checks It Performs:
| Problem | Description |
|---|---|
| Keyboard traps | Elements users can't tab away from |
| Broken link patterns | Links that don't work properly |
| Wrong click handlers | Interactive behavior on non-interactive elements |
| Bad tab-order values | Elements that override the natural reading sequence |
Concept 5: Visual AI Models — MolmoWeb and Molmo
The Big Idea: Instead of just reading code, PointCheck uses AI vision models to actually look at screenshots the way a human would.
Two Models, Two Jobs:
MolmoWeb
- Purpose-built for visual pointing on web pages
- Each screenshot is sent with the prompt: "Find the element that currently has keyboard focus"
- Returns the answer as a pixel coordinate (an exact location on screen)
Molmo
- Receives the same screen region
- Asked a direct question about what's visible — including whether a focus ring is actually visible on screen
- Confirms focus indicators are visually present, not just defined somewhere in a stylesheet
Why This Matters: Together, these models can catch focus indicators that exist in code but are invisible to the human eye — something no code-checker can do.
Concept 6: Additional Test Modules in PointCheck
The Big Idea: PointCheck runs several other visual and functional tests beyond keyboard navigation.
The Full Test Suite:
| Test | What It Checks |
|---|---|
| Keyboard navigation | Can users tab through the page? |
| Focus visibility | Are focus indicators actually visible? |
| Zoom at 200% | Does the page still work when enlarged? |
| Color-blindness simulation | Specifically tests for deuteranopia (most common form) |
| Contrast checks | Are text and background colors distinguishable? |
| Form submission | What happens when invalid data is entered? |
| Structural markup analysis | Is the page's underlying structure sound? |
Note: Keyboard navigation and focus visibility run as two separate tests because they evaluate different types of flaws.
Concept 7: Plain-Language Reporting with Olmo 3
The Big Idea: After all tests finish, Olmo 3 (a language model) writes a plain-English summary of the findings.
Why This Matters:
- Accessibility reports are often written for specialists
- Works wanted reports useful to product managers and designers — not just accessibility experts
- Olmo 3 was chosen specifically for its strength in plain-language synthesis
The Goal: Make the findings actionable for everyone involved in building a product, not just technical specialists.
Concept 8: Self-Hosting and Open-Source Models
The Big Idea: Works deliberately chose open-source models (Molmo, MolmoWeb, Olmo 3) so that PointCheck could be fully self-hostable.
What Self-Hostable Means:
- You run the tool on your own infrastructure
- Screenshots of your pages never leave your environment
- No data is sent to a third-party service for analysis
Why This Is Important:
- Privacy: Organizations don't have to expose their internal pages to outside APIs
- Control: Anyone can grab the model weights and run them independently
- Trust: The tool is transparent — you can inspect everything it does
Works' Own Words:
"You shouldn't have to send screenshots of your own pages to a third party to get an accessibility verdict."
Concept 9: The Perceptual Gap — What Only Vision Can Catch
The Big Idea: There is a category of accessibility problems that only visual inspection can find — things that require actually seeing the page.
Examples of Perceptual Problems:
- A focus ring that exists in code but is invisible against the background
- Text that technically meets contrast ratios but is still hard to read in context
- Layout that breaks at zoom levels in ways code can't predict
The Broader Significance:
- Traditional tools cover code compliance
- Human testers cover perceptual reality (but this doesn't scale)
- PointCheck aims to bridge that gap using AI vision — not replacing human testers, but extending what automated tools can reach
Concept 10: Open-Source as an Enabler of Unexpected Innovation
The Big Idea: Ai2 releases model weights, training data, and code openly because the most transformative applications are often ones they don't anticipate.
The Lesson from PointCheck:
- One person, working on weekends
- Using freely available open-source tools and models
- Built something that accessibility professionals said they wished had existed for years
The Broader Point: Open-source AI doesn't just help big organizations — it enables individuals to build targeted, meaningful tools for underserved communities that commercial products often overlook.
Summary: How All the Concepts Connect
User with disability
↓
Needs accessible website
↓
Traditional tools only check CODE → miss visual/perceptual problems
↓
PointCheck uses PLAYWRIGHT to simulate keyboard navigation
↓
Takes SCREENSHOTS at every step
↓
MOLMOWEB finds where focus is (pixel coordinates)
↓
MOLMO checks if focus indicator is actually VISIBLE
↓
Additional tests: zoom, color-blindness, contrast, forms, structure
↓
OLMO 3 writes a PLAIN-ENGLISH report
↓
All runs LOCALLY — no data sent to third parties
↓
Result: Accessibility problems that code-checkers MISS are now caught
The core innovation of PointCheck is simple but powerful: shift from checking what the code says to checking what the user actually sees. Every concept in the article builds toward that single idea.