SEO and accessibility

Judged, on the server#

POST /api/muse/audit asks a model the questions a pattern cannot settle: does the alt text describe what the picture shows, does a heading describe the section under it, would this link text still mean anything read out of context. The catalogue is server/muse/quality/audit-questions.js.

It is behind the Deep analysis checkbox because it spends tokens. The screen travels as data, in the user turn, under a header saying so (invariant Q5), and a verdict naming a question that was never asked is dropped rather than shown.

If it cannot run — no model, a timeout, an unparseable answer — the deterministic report is returned with a notice attached. Losing a complete, free report because a network call failed would be absurd (invariant Q1).

What the score does not mean#

Each dimension scores out of 100 and every one of them reports confidence: 'partial'. That is not modesty, it is arithmetic.

Source-only analysis knows the markup and nothing about the rendering. It cannot measure:

• colour contrast as painted,
• whether focus is visible,
• tap-target size,
• tab order as it actually runs,
• reflow at any viewport.

Those are half of WCAG. So nothing here claims a conformance level, and the panel prints the caveat beside the number rather than in a tooltip — a number whose caveat nobody reads is a number that gets quoted without it.

Two of those bullets have since gained a partial answer, and the report is careful to say how partial. Contrast is computed where an element paints both its own text colour and its own background, and a tap target's height is deduced from the classes that declare it. Neither is a measurement of a rendered page, which is why the colour family reports medium confidence and the tap-target family low, and why neither says anything at all about the elements it could not read.

A score of 100 means "nothing wrong in what this can see". It does not mean accessible.

Scoring counts distinct rules, never elements. Twenty images with no alt is one thing to fix; scoring it twenty times would drown every other finding and make the number meaningless. The panel still lists every offender.

The breakdown by family#

Under the two scores the panel lists eight families — headings, images, links and controls, forms, colour, typography, tap targets, structure — each with its own number. It is the same findings, cut a second way: seo and a11y answer how does this screen do, and the families answer where do I go and fix it. The two genuinely do not line up. An unnamed button and a missing alt are both accessibility errors and have nothing to do with each other, while img-alt and img-alt-redundant are one afternoon in one file.

The order is fixed in RULE_FAMILIES and deliberately not derived from the report. A list sorted by score reshuffles itself on every re-run, so the row someone was reading moves under their cursor between two clicks of Evaluate.

Each row carries a confidence, and it describes the method, not the run:

A row can carry no number at all, and that is the reason the section exists. A family with nothing to examine reads not applicable; one whose subjects the source does not describe well enough reads not measured. Neither reads 100. A screen with no form and a screen whose forms are perfect produce the same empty finding list, and only one of them has earned full marks — invariant Q4, applied one level below the panel's own caveat. The two words are kept apart because they are different facts about the screen: "there was nothing here to check", and "there was, and the source does not say enough to decide".

Judged findings are filed as well. Their rule ids exist only in server/muse/quality/audit-questions.js, so each question declares its own family there and the browser drops any name it cannot place — the same discipline as dropping a verdict that names a question nobody asked. Without it, a model finding about headings would take points off SEO while the headings row still read 100, and the breakdown would contradict the number directly above it.

The exported document#

A Mocky screen is one self-contained React component. It has no , no