Correctness is an architectural property, not a promise.
One standard governs every area of practice. Its governing idea is that a rule enforced by someone remembering to check has already been lost.
The grades
Rules are ranked by how hard they are to break, and the work of applying the standard is pushing each rule as far up that ranking as it will go.
The strongest form is a rule the system cannot express. Where a value must never be invented, the thing that would hold it does not exist — a component with nowhere to put a fabricated number cannot display one. Nothing has to be remembered, because the mistake has no representation.
Below that is a rule the system refuses. The work stops before it ships, with the reason and the location, at the moment somebody wrote the problem rather than months later when a reader finds it.
Below that is a rule the system measures. A machine reports a number, so that a disagreement about whether something looks wrong becomes a disagreement about a measurement, which is a considerably shorter conversation.
Weakest is a rule a person is asked to observe. Any rule that begins "remember to" has already been identified as one that will eventually be forgotten, and it is written down as a candidate for promotion rather than treated as finished.
What gets enforced
Claims and the conditions attached to them are inseparable. Where a statement carries a legal requirement, the two are one thing, so a page cannot show the claim while losing the condition through an ordinary edit.
Anything the business cannot substantiate has no way to be stated. There is no facility for a testimonial, a rating, a review score, or an outcome figure without a source — not a policy against them, an absence of the means.
Every published word is subject to automatic checking before it reaches a reader, including the words assembled from several pieces at the moment a page is produced. Checking what was written and not what was published is the gap that lets composed claims through.
Accessibility is measured against the values actually in use and recorded alongside them, so that a color change which quietly breaks a contrast requirement is caught when it is made.
Where the firm declares how much of something exists — how many pages, how many areas — the number is asserted and checked. A count that is merely observed can never disagree with itself, and so can never catch anything.
How work is verified
Verification runs against what was actually produced, not against what was intended. The distinction sounds academic and is the source of most of the defects this standard exists to prevent.
Checks are calibrated before they are trusted. Each one is required to demonstrate that it catches a known-bad example and stays silent on a known-good one, because a check nobody has proven can fire is indistinguishable from one that does nothing, and a check that fires on correct work is worse than no check at all — it teaches everybody to ignore it.
A false alarm costs more than a missed defect. Two over-broad checks once produced several hundred findings that were not real, more than every genuine defect found across five projects put together, and the lasting damage was that the next real finding arrived in a channel nobody trusted.
Release is confirmed against the live address rather than against a report of success. A completed build and a working site are different claims, and the most common incident is a successful release of the wrong thing.
Everything above runs on every change. A check that has to be invoked by hand is a check that runs when somebody remembers, which is the failure the standard began by rejecting.
What the standard costs
It is slower at the start. Deciding scope in writing, arguing the case for each addition, and building the checks before the content they judge all happen before anything visible exists, and that is uncomfortable for everyone involved.
It constrains what can be offered. Some entirely ordinary requests cannot be met inside it, and the honest response is to say so rather than to make an exception quietly and let the standard become decorative.
It is not free to maintain. Checks need calibration, and one that has quietly stopped firing is worse than none, because everybody believes it is working.
What it buys is that a specific class of failure stops being possible instead of becoming less likely, and that the difference is demonstrable rather than asserted.
What it does not claim
It does not make anything correct. It makes particular, well-understood ways of being wrong unrepresentable, which is a narrower promise and one that can actually be kept.
It is not a certification, it has not been audited by anyone, and no external body has assessed it. It is a description of practice, published so that it can be read, checked, and argued with — which is the whole of what distinguishes a standard from a claim about one.