How axe-core keeps false alarms rare enough to trust

How does an automated checker stay right often enough that people stop second-guessing it?

axe-core is the accessibility checker behind many browsers’ developer tools. It’s built to avoid false alarms, and when it can’t tell, it says so instead of guessing. I’ll study how its rules are designed and tested, then spend a week building a way to measure how often a design rule is right.

Format: Teardown
What I’ll build: A test bench that runs each design rule against labelled examples, reports how often it’s right, and keeps a rule out of the default set until it clears a bar.
What I’ll reuse: A precision bench for design rules, so only rules that are usually right get to block anything.
Used by: kbui
Status: Planned · Nov 2026

What I’ll look at

How axe-core decides a rule is reliable enough to ship, and how it tests each one against pages that should pass and pages that should fail. When it reports “needs review” instead of a violation. And how it handles what a rule can’t see, like text contrast on top of an image.

My guess

I expect axe-core is trusted mostly because it’s willing to say “I don’t know”. A rule with three answers (pass, fail and check by hand) can afford to be strict about the first two.

How I’ll prove it

I’ll build a set of small screens that follow or break each rule, label them by hand, and run every rule against them.

A rule that’s wrong too often drops to “check by hand” until it improves, so the checker never blocks on a guess.