History
Timeline
Here you'll find every version and milestone of the project in chronological order, from the idea to today. Be part of this story.
- Milestone
The discomfort that became a project
At the “Who's left out when AI works?” roundtable in the TDC Summit São Paulo Community Lounge, accessibility keeps being framed as compliance and punishment. The discomfort leaves the room as a question: what if the rules entered the AI's context before its first answer?
View TDC's Community Lounge (opens in a new tab) - Milestone
The repository goes public
The premise from the first commit: accessibility as a precondition, not a retrofit — a markdown file any coding agent can read.
View the repository (opens in a new tab) - Milestone
First public mention — on the same day
Frontend Dogma, Jens Oliver Meiert's accessibility archive, lists the project on the day it is born.
View the archive (opens in a new tab) - Milestone
First star, one day later
The first person stars the repository in under 24 hours (@mewmewdevart, recorded by the GitHub API).
- Milestone
First fork
On day two, someone takes a copy to build on (@Hiltonjunior). A standard only exists once it leaves its author's hands.
- Milestone
“AI doesn't fix inaccessible systems. It scales what it finds.”
Vitor David (ABNT CB-040, the committee behind Brazil's NBR 17225) publishes the sentence that sums up the problem this project attacks — still quoted on this site today.
View the post (opens in a new tab) - Milestone
First independent test
Maria Eduarda Iwashita runs a blind-baseline A/B test against a pre-1.0 version — without asking permission, which is how science works. Her findings return to this timeline in August.
View the study (opens in a new tab) - Milestone
The first talk in the project's name
At the Design Imparável meetup (Sem Parar + IxDF São Paulo), on a night of heavy rain, the project is presented publicly for the first time — from the TDC roundtable to the thesis that accessibility is decided before any prompt.
Watch the talk (opens in a new tab) - Release
v1.0.0 — profiles, contract, library
The three compliance profiles (Shield, Standard, Launchpad), the AI behavior contract and the technical guide library form the first stable version.
Release notes (opens in a new tab) - Milestone
Selected for Claude for Open Source
The project joins Anthropic's program supporting open-source maintainers worldwide.
- Release
v1.1.0 — normative precision and native translation
Every requirement now cites its WCAG criterion by number, separated from House Rules†. And the translation layer for iOS, Android, React Native and Flutter — the only one in the field — joins the library.
Release notes (opens in a new tab) - Milestone
This site goes live — built under its own standard
Dogfooding at the most demanding profile (Shield), with a public conformance report: the site follows the standard it promotes.
View this page's code (opens in a new tab) - Milestone
The standard reaches a real building
The São Paulo CEUs portal (Faculdade Phorte × City Hall, design by Marcelo Vignola) is born with A11Y.md in the AI's working rules — and the conversation spills off the screen: the team starts studying physical accessibility improvements in the educational centers themselves.
View the design case (opens in a new tab) - Milestone
Benchmark pre-registered — before any data
Methodology and prompts published before collection, so the result can be challenged and reproduced. The standard demands evidence from everyone else; it holds itself to the same bar.
View the protocol (opens in a new tab) - Release
v1.2.0 — evidence gets a trigger
The evidence artifacts stop depending on someone remembering to ask: report, decisions and exceptions now have a trigger in the contract, and the repository gets its first executable checker.
Release notes (opens in a new tab) - Release
v1.3.0 — a community question becomes a rule
“How does the standard handle alt text when the user supplies the image?” It didn't. Now it does: Image Evidence, with the human decision as part of the rule, not a courtesy.
Release notes (opens in a new tab) - Release
v1.5.0 — time-based media and cognitive accessibility
Video and audio get a contract of their own, and the cognitive criteria of WCAG 2.2 enter the standard's floor.
Release notes (opens in a new tab) - Release
v1.6.0 — the independent test becomes a rule
The bug the June study found — an orphaned ARIA reference that passed every tool — enters Section 6 as an anti-pattern, with named credit to its finder.
Release notes (opens in a new tab) - Release
v1.7.0 — the generator no longer signs alone
Independent Verification joins the contract: the agent that wrote the code cannot be the sole witness that it conforms. Five new guides — including generative interfaces — and voice-control validation in the checklist.
Release notes (opens in a new tab) - Release
v1.8.0 — the core slims, the library deepens
The core file loses 10% without losing one obligation: rationale moves into the guides, where lazy loading serves it on demand. A new check keeps it that way — a placement rule, never a content cap.
Release notes (opens in a new tab) - Milestone
The community gets doors
Discussions opened, contribution forms (field reports, errata, proposals, gallery) and a complete security policy. That day, the repository stood at 363 stars and 15 forks — with 18 contract rules and 29 guides in the library.
Join the conversation (opens in a new tab) - Milestone
The two efficacy studies are published
Two preregistered studies with a placebo control and blinded judgment: 400 generations and 30 real-agent journeys. The full account — wins, ties, and four documented instrument defects — becomes a permanent page on this site.
Read the full study - Milestone
The conversation reaches Natura
An accessibility and inclusion panel inside Natura, alongside Simone Freire, founder of the Web para Todos movement, and Vitor David, of ABNT’s CB-040 committee and author of the sentence that opens this site. The subject stepping out of the specialists' circle and into organizational decisions and culture — precisely this project's bet.
See the panel recap (opens in a new tab) - Release
v2.0.0 — the post-benchmark milestone
Every major change traces to a number in the studies: contrast computed by deterministic tooling, the Brazilian NBR 17225 norm mapped, a pioneer guide for automatic Libras translators, the agentic-web layer — and an evidence-gated token diet that paid for all the growth.
Release notes (opens in a new tab) - Release
v2.0.1 — the Study 3 consequence
Study 3's hardest criterion failed, and the consequence table written before collection collected on the promise: Section 6 now names the exact mold that survived both versions — the carousel announced as tabs without the tab roles — with its two honest exits. Nothing else changed.
Release notes (opens in a new tab) - Milestone
Study 3 — the new version returns to the bench
The second formal movement: 2.0.0 against 1.8.0, on the same agents, publicly registered before the first data point. Both tie at the automatic ruler's ceiling, the hardest criterion fails on both sides — and the measurement that does not tie is accountability: 19 of 19 with the standard, 0 of 20 without.
Read the full report - Release
v2.0.2 — the addresses an agent fetches
No obligation changed. What changed is the address of everything an agent fetches: the invocation phrase, the lazily loaded guides and the scripts now point at the plain file on GitHub, not at the page. The phrase also stops naming "A11Y.md" before the link, because .md is an internet domain and an IDE chat could turn the name into a link to the wrong site.
Release notes (opens in a new tab) - Release
v2.1.0 — the receipt
The studies measured what the model does when it reads the standard. Now the standard also holds the evidence the agent writes afterwards to account: the checker runs before every delivery and says so when it is denied, reads in source what axe cannot see, and every contrast ratio in the report is recomputed by the formula instead of believed.
Release notes (opens in a new tab) - Release
v2.2.0 — the scenarios
Since v1.7.0 someone who did not write the code has to witness that the interface works, but nothing told that person what to look for. Now every component guide ends with verification scenarios: what must happen by keyboard and what the screen reader must announce, on desktop and on a phone. The shape came from Magoo, Marcelo Sales' behavior library. Also in: a guide for checkbox, radio, switch, slider and select, and the field that says which version of the standard a report was verified against.
Release notes (opens in a new tab) - Milestone
The conversation reaches the Web para Todos Volunteer League
For the first time the project is presented to the accessibility community, not to the design or the technology one. The 28th meeting of the Web para Todos Volunteer League hosts the conversation, led by Simone Freire, founder of the movement, about how the standard was born, what it does and how to use it. There are blind people in the room, and the conversation opens with a self-description.
Meet the Volunteer League (opens in a new tab)


