can people use what an llm builds?

ask a model for a web page and you get one in seconds. it renders and it looks right. the harder question is whether a screen-reader user, or just a first-time visitor, can get anything done on it.

two lenses

my research asks that question in two parts. accessibility is the one with a checklist: wcag spells out contrast ratios, names for controls, keyboard operation and so on, and tools like axe can test a lot of it automatically. usability is the fuzzier one. nielsen’s ten heuristics cover things like visibility of system status, consistency, and error prevention, and they need judgment, not a linter.

each lens misses what the other sees. a page can pass every automated rule and still make you guess what a button does, or read clearly and still trap a keyboard user. so the project scores both on the same page.

the loop

people rarely take the first output. they ask the model to fix things, over and over. so the thing to measure is the loop: generate a page, evaluate it, ask the model to repair what the evaluation found, and repeat.

that raises questions i don’t know the answers to yet. does repairing accessibility make usability worse, or the reverse? do the fixes hold over several rounds, or does the model undo its own work? does it matter who gives the feedback, a tool, a checklist, or a person?

what it looks like by hand

i ran both lenses on this site. axe passed structure, names and landmarks, and flagged only contrast. measuring contrast myself showed most of my secondary greys failing. a manual pass showed my single-key shortcuts breaking a level-a wcag rule that no tool reported. the usability pass found links that didn’t look like links. each method caught something the others missed, which is the reason to use more than one. here is the contrast result, before and after the fix:

where it stands

the experiment runner is still being built, so there are no results to report, and i’d rather say that than guess. the planning, the research documents and the wcag reference material are on github. i’ll post what the loop shows when it shows something.