Component proving ground

Outcome — the only coloured element

Two vocabularies at two grains, one visual axis. Register density is the word alone in the outcome colour; detail density is a chip. Two of the five carry no hue at all, because the absence of a conclusion does not belong on a good-to-bad axis.

Severity — no hue, ever

Outcome and severity are orthogonal fields on the same finding, and only one of them can be what colour means. Weight steps up at high and critical; nothing else does.

Figures and timestamps

Fractions stay fractions. A timestamp states its zone before hydration and switches to the reader's on mount, so it is never a time in an unstated zone.

The four things a finished item can say with no deficiency

All four carry status === "analyzed". Read them as shapes before reading the words: a dashed panel waiting for a person, a heavy rule at full height twice over, and one result that is allowed to look settled.

Failure, on its own

Not red. Red is a dropped connection or a failed submit, and it never appears beside a conclusion — including where a conclusion is missing.

Paper

The application is the desk; the workpaper is paper on it. Warm white against cool grey, square corners, a real margin, a 72ch measure, and serif — because serif means this text is going into the deliverable.

Box registration

The claim this product rests on is that a conclusion points at specific pixels. These boxes are positioned in percentages of the rendered image, so they need no resize listener and survive zoom, device-pixel-ratio changes and column reflow — and the container they sit in is exactly the image's box, with no border or padding, because either would offset every box by its own width. Boxes rest as hairlines and take the accent only when pointed at, so the evidence underneath stays readable. E arrived with its coordinates inverted; it is normalized, drawn dashed, and not silently dropped.

Resize the window or zoom the browser: every box should stay on its block at any size, because nothing here is measured in pixels.

• 1A — upper lefty 100–300 · x 50–250
• 2B — lower righty 600–900 · x 700–950
• 3C — centre, wide and shorty 400–500 · x 400–600
• 4D — flush to the top-left cornery 0–60 · x 0–40
• 5E — arrived invertedy 700–500 · x 300–200

Did it run?

Colour in this product means outcome, and none of these six stages is an outcome — so state is carried by the left rule instead: 2px solid for a stage that ran, 4px (the Notice weight) for one that failed, 2px dashed and faded for one that did not run at all. Every stage renders in every state. A pipeline that hides its skipped stages is not a shorter page, it is a page that lies about being complete. Two stages are marked no model, because which parts were not the AI is a question an auditor has to answer to their own reviewer.

Register density, at 200 rows

The scale the register is actually used at, rendered through the shipped component rather than a mock. Rows are monochrome: exactly one element carries a hue — the conclusion word — and severity, which is the other axis a reviewer sorts on, gets none at all. The adoption rule under each conclusion is the finding byline's mechanism at 36px: solid where a human stands behind it, dashed where nobody does, absent where there is nothing to adopt. Control refs come from three frameworks with different shapes, titles run from 9 to 118 characters, and the counts include four digits, because a column that only aligns for the data you had is not aligned.

A conclusion, and the four ways of standing behind it

The terminal decision. The design problem is not presenting a conclusion — it is presenting one a machine reached without it acquiring the authority of one a person reached. The byline is the mechanism: an unadopted finding carries a drawn, empty signature line, the same shape a paper workpaper uses, so it costs nothing from the rationed accent and survives a photocopy and a screen reader intact. An override is shown as a change, never as a value — a severity a reviewer raised is a different artefact from one the model produced alone, and the two are identical if you print only the effective value. Confidence is the model's own number and is labelled as such; there is no bar, no quality word, and no tick.

The workpaper, on the desk

The deliverable, and the one screen where the metaphor is doing literal work: the application is the desk — cool, grey, fluid — and this is a sheet of paper lying on it. Warm white, square corners, a real margin, a 72ch measure, serif. Sections are rendered by the shape of their content rather than by their name, so a section type this build has never seen still prints, with a generated heading and a note saying the heading is ours. A frontend that renders only what it was taught would silently print a shorter workpaper than the one on record. Empty sections say they are empty rather than vanishing — `management_response` is blank by design until a person fills it, and a workpaper that hides its blanks looks finished.

Ink, rationed