The conformance note, signed
A note listing only what passes is marketing; the one worth signing names the criteria the component fails, by number, with the remediation cost of each — which is what lets a portfolio company argue about priority instead of about facts.
The note is full. There is a scope, a target version and a level, an accessibility-supported line, findings with numbers and levels and costs, a provenance column, a list of what is outside automated reach, and a list of what nobody checked. All that is left is to put a name and a date on it and give it to somebody.
Which is the moment the temptation arrives, because a document that opens with a column of failures reads like a confession, and there is an obvious edit available: lead with what passes, move the failures into an appendix, and let the summary line say the component is broadly conformant with known gaps.
Do not make that edit. This lesson is about why.
Where people get burned
Everything in this lesson is this course’s position, and none of it is citable. No W3C document defines what a component-level conformance note is, or says that listing failures is the credible form of one. The WCAG 2.1 and 2.2 conformance requirements, the Understanding Conformance document and the ARIA Authoring Practices Guide were all searched for a treatment of remediation cost, component-level notes, or what makes a claim credible, and none of them addresses any of it. What follows is an editorial argument in the course’s own voice. Attributing it to the W3C would be a fabricated citation, which in a course about citing accurately would be the only unforgivable error available.
The one thing the specification does give you
Understanding Conformance contains two sentences that support the spirit of the argument without making it. The first is that no conformance claim is required in order to conform. A claim is an optional artifact. Nothing obliges you to publish one, which means the only reason to publish one is that it does something for the reader — and a document that exists to reassure rather than to inform is doing nothing for them.
The second is stronger: content can satisfy every success criterion and still not be usable by people with a wide variety of disabilities. The Working Group is saying, in its own informative material, that full conformance is not a guarantee of usability. If passing everything is not a guarantee, then a document listing only passes is guaranteeing something the specification explicitly declines to guarantee.
That is as far as the sourced argument goes. It gets you to “conformance claims are optional and passing is not sufficient.” It does not get you to failures, costs, or signatures. Those are next, and they are the course talking.
Why the failures are the useful part
The argument is about what kind of conversation the document makes possible.
A note listing only what passes produces one of two responses. Either the reader believes it, in which case nothing gets fixed and you have spent your credibility buying silence, or the reader does not believe it and now has to establish the facts themselves, in which case the document wasted their time and yours. Neither response schedules any work.
A note listing failures by number, with a level and a cost on each, produces a different conversation, and the difference is the whole point: the argument moves off the facts and onto the priority. Nobody in the room is disputing whether the trace panel is keyboard operable, because there is a criterion number, a level, and a test anybody can re-run. The only remaining question is which rows get done this quarter. That question is answerable. “Is this actually a problem?” is not, and it is the question a passes-only note guarantees you will be asked.
The costs are what make it a schedule rather than a complaint. A finding without a cost is an opinion — someone else has to estimate it before anything can happen, and that someone is usually the person least motivated to do it. A finding with a cost is a line item. It can be traded against other line items, which is the only mechanism by which anything ever gets prioritised in a company.
The rules this note is written under
Each of these has an earlier lesson behind it, and each is the course stating a standard rather than reporting one:
- It is a note, not a claim. Conformance is defined for full pages and complete processes, so a component cannot conform. The title of the document has to say what it actually is, or the first reader will quote it as something it is not.
- Every row carries number, exact name, level, test, verdict, cost and provenance. The exact name matters because a wrong name gets a whole finding dismissed as sloppy. The level matters because it is how a reader sorts. The test matters because it is what makes the verdict reproducible by somebody who does not trust you.
- Never mark a row that no tool covers as tool-checked. There is no axe rule for live regions, so every 4.1.3 row in this note is
humanorunchecked. A falsetoolis worse than an empty table, because it survives re-reading. - The accessibility-supported line is not aspirational. If no browser and screen reader combination has been verified, it says “none verified.” An empty list is a finding. An invented one poisons everything below it.
- Name the states you never entered. The error state, the mid-run state, the open-dialog state. If the checked-by-nobody list comes out empty, it is wrong.
- Mark the reasoning that is yours. The streaming pattern, the interrupt rule, and the application of 2.2.2 to an agent run are all constructions rather than citations, and each says so in the note. This is the rule that gets edited out during tidying, and it is the one that makes the rest trustworthy.
- Say what you could not read. This course cites EN 301 549 nowhere, because ETSI’s published PDF returned HTTP 403 and every readily available description of it is written by a vendor selling remediation against it. The axe-core coverage figure is quoted from the repository without a version number, because npmjs.com returned 403. Both absences are in the note. An honest gap is a finding; a gap papered over with a second-hand summary is a liability.
What the signature means
The date is the date of the assessment, not the date the component shipped, because the note describes a build and builds change. The name is a person, not a team, because a document nobody signed is a document nobody is answerable for.
And the last line of the note is the sentence the tooling lesson asked for: what a green CI run establishes for this component, phrased so you would be comfortable reading it aloud to a portfolio company’s engineering lead. Something in the shape of: the automated suite evaluates a subset of Level A and AA criteria against the default rule set for one version of the guidelines, in one page state, and produces no verdict at all on status messages, focus order, error suggestion, or anything requiring a screen reader. That sentence is the honest product of this entire course.
Check your recall
Answer from memory — no scrolling back.
Retrieval check
A portfolio company’s engineering lead reads your note and says: “This is a list of everything wrong with our component. Why would we let you publish it?” Answer in three sentences.
Check your answer
It is not for publishing, it is for scheduling — every row has a criterion number, a level and a cost, so you can decide what goes in this quarter instead of arguing about whether any of it is real. The alternative document, the one that lists what passes, gets you a procurement question you cannot answer, because a note that only claims passes is a note somebody will eventually test.
And start with the Level A rows, whatever they turn out to number, because A is the bar underneath every conformance level there is — they are the least arguable work on your roadmap this year.
The move in that answer is refusing the frame. The note is not an accusation and it is not a marketing asset. It is an input to planning, and it is worth exactly as much as its worst row is honest.
Hands on
Sign the note
Done when: CONFORMANCE.md is complete and signed: the header block has no placeholder text left, every finding row has all seven fields, no row that no tool covers is marked tool, the accessibility-supported line is either verified combinations or the words “none verified”, the constructed recommendations are labelled as constructions, the unreadable sources are named, and there is a person and an assessment date at the top.
- Read the note top to bottom as a stranger would, in one pass, without editing. You are looking for the sentence that would make a hostile reader stop trusting the rest — usually an adjective where a number belongs, or a verdict with no test behind it.
- Audit the provenance column against reality one last time. Every
toolrow must correspond to a rule that actually exists and actually ran. Every 4.1.3 row ishumanorunchecked. - Check the accessibility-supported line. If nothing has been verified against a named browser and screen reader pair, it says “none verified” and that is a correct, finished answer — not a gap to fill in before signing.
- Find every recommendation that came from this module and confirm each one is labelled as your construction: the streaming pattern, the interrupt rule, and the application of 2.2.2 to an agent run. Confirm the mid-run discoverability finding still carries no criterion number.
- Add the sources you could not read, by name and by reason. EN 301 549 at HTTP 403, the axe-core version at HTTP 403. Two lines, and they buy more trust than any passing row in the table.
- Write the final line: what a green CI run establishes for this component. Then put a name and an assessment date at the top, and bring the whole note into the chat. I will read it the way an auditor would — looking for a row I cannot reproduce, a claim with no source and no label, and a checked-by-nobody list short enough to be aspirational.
What this does not cover
This is the last lesson, so the boundaries here are the course’s rather than a sibling’s. The criteria themselves are not something to remember from a lesson you read once — the criteria reference page is the part of this course meant to be reopened, and it is where the numbers, names, levels and plain-English tests live for the failures that actually recur in component work.
Four things this course deliberately never covered, so you do not go looking for them here. Document accessibility — PDF and Office — and native mobile are different technical standards with different techniques. Overlays and automated-remediation products appear only as something to be sceptical of. Legal advice is out entirely: the course names which regulation cites which version of the guidelines and stops there. And EN 301 549, the European harmonised standard, is absent because ETSI’s own PDF could not be read and the alternatives were all written by vendors selling remediation against it. If a portfolio company needs it, get it from ETSI directly.
The one claim in this module with a shelf life is the empty shelf itself. The W3C’s generative-AI accessibility draft was silent on streaming announcement as of 5 September 2026. That will change, and the announcing-the-stream lesson is written so that when it does, the fix is a re-read and a date rather than a rewrite.
Read this next — primary source
Understanding ConformanceW3C Web Accessibility Initiative — free, informative (non-normative)
This is the one document with any bearing on what follows, and it is worth re-reading now that you have findings rather than theory. Two sentences do the work: that no conformance claim is required in order to conform, and that content can satisfy every success criterion and still not be usable by people with a wide variety of disabilities. Read past them into the optional-components section on what a claim must contain if you choose to make one, because everything this lesson argues is an attempt to build the honest analogue of that for a component, which the W3C does not define.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.