Why presence telemetry is not a design score

We had a page that passed every anti-slop rule and still looked like nothing. The first experiment tried to measure the opposite: not the absence of defects, but visible commitment. It produced a neat number and an important failure.

The initial hypothesis

Three rendered signals looked promising:

• Display scale — largest included leaf-text size divided by body size.
• Surface coverage — viewport area visibly departing from the page paper.
• Container density — framed or shadowed elements, with one- and two-sided separators counted separately.

The first implementation summed normalized versions into a 0–100 score. On a small set it appeared to reproduce a human ranking. Widening the control set invalidated that conclusion.

The score rewarded exactly the wrong things

A purple gradient covers a lot of area. An oversized headline creates a large type ratio. A uniform card grid creates many boxes. Those can all be deliberate, and they can all be generic slop. The number could not tell the difference.

It also mixed scopes: type and boxes came from the fold while colour came from the whole document. That made the denominator change with page length and turned unlike surfaces into false comparisons.

The correct conclusion is not “find better weights”. It is that these observations do not form a quality dimension.

The corrected contract

pnpm scene:audit now emits raw, conditioned telemetry:

• Every geometric signal uses the same initial-viewport scope.
• Surface and container coverage use exact axis-aligned unions, so overlaps count once.
• Solid colours are alpha-composited over ancestor paper.
• background-image is reported separately as an upper bound because CSS geometry cannot reveal transparent pixels inside a photograph or gradient.
• Unsupported colour spaces are counted as uncertainty rather than silently becoming black or transparent.
• The type census states its population: rendered leaf text in the viewport, excluding decorative nodes, fixed/sticky site chrome, the comparison switcher and explicit audit exclusions.
• Each scene carries a surface profile. Marketing, reference, authentication, commerce and dense-product surfaces have different valid shapes.

There is no weighting, ranking, “good” band or pass/fail threshold.

What the gate actually protects

CI does not fail because a page has “too little commitment”. The local build runs the pure geometry/colour tests. During bootstrap, a separate pinned-Chromium job renders the fixed scene corpus and uploads a validated telemetry artifact for review. Once that artifact is intentionally committed as the baseline, the job can switch to byte-stable drift checks so a change fails only when measurement or scene output moved without the artifact being reviewed.

This is a reproducibility gate, not a taste gate.

What remains useful

The signals are still useful during review. A sudden jump in container union area may reveal that a product screen became card soup. A changed type census may reveal hidden chrome entering the sample. A background-image upper bound can explain why solid colour coverage fell. But interpretation begins with the surface type and ends with the rendered page.

The failed score was the useful result: measurable does not automatically mean rankable.