AHD · Glossary

The vocabulary.

About twenty words on this site carry a narrower meaning than usual. They are defined below, alphabetically. Pages link here on first use.

Brief

A short structured file that states what a page is for, who it is for, what it must contain and what it must avoid.

One brief drives every model in a run, so the only thing that varies is the model and the condition. The brief used in the published runs is shown field by field on the methodology page.

Cell

One model under one condition.

Ten models measured in two conditions makes twenty cells. A cell is the unit every published number attaches to, because a model measured raw and the same model measured compiled are two different measurements.

Compiled (condition)

The model receives the same brief, plus an AHD system prompt built from a style token.

The system prompt names the style direction, lists the patterns to avoid and asks the model to cite the rule behind each decision. Compare raw.

Delta also: reduction, regression

The change in average tells per page between the raw and compiled measurement of one model.

A fall in tells is a reduction, written here as a percentage with a down arrow. A rise is a regression, written with an up arrow. A regression means the AHD layer made that model's output worse, and those are published in full.

Lint

To check finished output against a fixed list of rules automatically, the way ESLint checks code.

AHD lints rendered HTML and CSS rather than source code, so it works whatever tool generated the page.

Manifest

The record a run writes describing exactly what it did.

It holds the brief, the style token, the exact model identifiers, the number of samples and the timestamp. Without it a published number is only an assertion.

n

How many times each cell was sampled.

n=30 means every model was asked thirty times under each condition. Larger n narrows the uncertainty around the result. See Wilson interval.

Pre-fix and post-fix

Two readings of the 24 April run: before and after token-aware linting, which lets a token switch off the rules it deliberately rejects.

The samples are identical in both readings. Only the rules that scored them changed. Both readings are published so the correction itself is visible.

Raw (condition)

The model receives the brief and nothing else. No style token, no AHD system prompt.

This is the baseline: what the model produces when asked for a page with no design direction from AHD. Compare compiled.

Rule

One automated check that fires when it finds a specific pattern.

AHD ships 38 rules that read HTML, CSS and SVG, 14 that read a screenshot through a vision critic and 6 that check a page at phone width. Some tells need more than one rule to cover them, which is why there are more rules than tells.

Sample

One page produced by one model on one attempt.

Every published run reports how many samples it attempted, how many failed and how many reached the linter. A wide gap between attempted and scored says the model struggled with the instruction.

Serving path

The specific host, release and configuration a model was reached through.

The same model weights served by two different hosts behave differently, because quantization, chat template and default settings belong to the host. Every cell on this site names its serving path for that reason.

Slop

Design that is recognisably machine-generated because it repeats the same small set of defaults.

The purple-to-blue hero gradient, three equal feature cards, one border radius on everything. The word describes a distribution, not a verdict on quality.

Tell

One named, specific pattern that marks a design as AI-generated.

AHD names 39 of them across web, graphic and typographic surfaces. The full list is the taxonomy.

Tells per page

The average number of rules that fired across the samples in one cell.

It is a proxy for how strong a page's machine fingerprint is, not a score for how good the design is. A page with almost nothing on it fires almost nothing. Read it next to the rendered sample, never on its own.

Style token

A named design direction in a form you can hand to a model.

Each one declares its grid, typography, palette, the patterns it forbids and the quirks it requires. AHD ships ten. swiss-editorial and post-digital-green are the two used in the published runs.

Vision critic

The part of AHD that judges a screenshot rather than the code behind it.

Some tells only exist once a page is drawn, so no amount of reading the HTML will find them. Those rules run against rendered pixels.

Wilson interval

The band of uncertainty around a measured proportion, such as the share of samples in which a rule fired.

At 5 samples per cell the band is roughly plus or minus 35 points, which shows a direction but does not support quoting a number. At 30 samples it tightens to roughly plus or minus 18 points. A result inside the band is indistinguishable from zero. The headline reductions are a different quantity, a change in mean tells per page, and the interval is used there as a rough guide to resolution rather than an exact bound.

Adjacent reading: how we measure, the taxonomy, every run.