Aug 26, 20265 min read

a11y Companion v3.0.0: A Checkmark Isn't Evidence

v2 argued that a contrast tool should refuse to claim what it cannot evaluate. v3 makes the same argument about checklists, and gives every checked box a way to be proven.

a11yaccessibilitywcagfigmafigjamopen-source

Share on X Share on LinkedIn

♿️ a11y Companion v3.0.0 is live.

The checklist is still the checklist. What changed is that the widget can now tell you whether the file actually earned the checkmarks.

Source checklist: The A11Y Project. Standards target: W3C WCAG.

The gap I kept running into

v2 was about a contrast tool refusing to claim what it cannot evaluate. Shipping it exposed the next problem, one floor down.

A checked box is a claim, not a measurement.

Nothing in the widget knew whether the file met the criterion sitting next to the checkmark. A team could complete an entire checklist against a file that failed contrast, and the widget would show a full green ring. The checklist was honest about what it asked. It had no way to be honest about the answer.

So v3 asks a different question. Not "did someone check this box," but "what would it take to prove it."

Three things a claim needs to become evidence

It has to reach the people who act on it

Findings that live inside a widget are findings nobody fixes. v3 promotes audit results to Figma's own Dev Mode annotations, with the WCAG Understanding page linked and the failing property named: width, height, fill, font size, line height.

It reads and rewrites the annotation array rather than replacing it, so annotations you wrote by hand survive, and running the same audit twice does not produce two copies.

Audit findings promoted to native Dev Mode annotations

It has to expire

A sign-off that outlives the design it approved is worse than no sign-off, because it looks like diligence.

When you sign a section in v3, the signature records what it actually reviewed: item ids, completion state, and the checklist revision. Change items underneath a signed section and the stamp turns amber. The approval does not quietly stay valid.

Section sign-off with reviewer stamp and stale-state warning

It has to leave the file

A review that only exists in one Figma document stops being useful the moment anyone needs it elsewhere.

v3 compiles a Canvas Record, one Figma frame holding progress, sign-offs, audit summaries, contrast evidence, and priority findings. Re-running it refreshes the same frame instead of littering the page with copies. The same review exports as Markdown, CSV, or JSON.

Report export as Markdown, CSV, and JSON

What's new in v3.0.0

Three on-canvas audits: Targets, Text, and Tokens, run from under the progress ring
Live contrast watching that re-checks the selection as it changes, with shared ownership so the team sees who is driving
Color vision preview for protanopia, deuteranopia, and tritanopia
WCAG level filters, including a "New in 2.2" filter for the nine criteria WCAG 2.2 added
Native Dev Mode annotations, individually or in bulk
Canvas Record, exportable as SVG or PDF
Evidence-bound section sign-off
Report export as Markdown, CSV, or JSON
Twenty locales, up from thirteen

Where it still says no

The honesty argument only holds if the tool applies it to itself. Three places v3 refuses to overstate:

Audits share a capped slot. Past the cap, the counts stay truthful rather than reporting only what fits on screen, and the Canvas Record will not call itself ready on capped evidence.
Reports disclose when their rows are representative rather than complete. A grouped root cause is labeled as one.
The color vision preview is a simulation, not a second grade. WCAG contrast stays the normative result. Three more pass/fail badges would feel more thorough and mean less.

Contrast thresholds are also evaluated at full precision, so a pair sitting just under the line cannot round its way into a pass.

I ran it on itself

v3 had to pass its own checklist. That caught real problems: checkboxes that failed WCAG 1.4.11 non-text contrast, severity conveyed by color alone, and WCAG references that rendered wrong in dark mode.

Testing the token audit against Vercel's Geist file caught another one. My pairing logic looked for names containing "text" and "bg", which found roughly twelve pairs in a system named gray/100 through gray/1000. It now infers role from the trailing numeric step instead.

Being wrong in a way you can measure is the whole point.

Why I keep tightening this

Accessibility is personal for me. I have been the eyes and ears for my legally blind and deaf brother for decades. I have watched him burn time on sites that do not care.

Compliance language never helped him. Working designs did. That is why v3 went after evidence instead of more checkboxes.