Three successful clicks. One tab stop.

Results

The table can be scrolled horizontally.

The original files are preserved. The contrast counter-test changes only the notebook background (#687b40 becomes #647640); the cursor counter-test starts from the original and adds .command { cursor: default; }. The two changes are separate.

MCP calls

browser_snapshot supplies the lines below; browser_click then uses the returned reference. The counter changes from 0 to 1 for every variant. References are local to each session. A script selects the target: no autonomous model decides the action.

A
- generic [ref=e14] [cursor=pointer]: Ajouter au panier

B
- button "Ajouter au panier" [ref=e14] [cursor=pointer]

C
- button "Ajouter au panier" [ref=e14] [cursor=pointer]

With cursor: default, the [cursor=pointer] annotation disappears; the reference and working click remain for all three variants.

Keyboard tests

The test sends two Tab presses in a fresh page, requires the button to have focus, then presses Enter or Space and expects a count of 1. Two tests pass for C; two fail for A and two for B, in each of the two datasets. Reports preserve all eight failures and four passes. The A/B test processes exit with code 1.

keyboard.spec.cjs · playwright.config.cjs

A failed test can be an expected experimental finding. It remains a failure of the required behaviour.

Lighthouse scores

The contrast warning concerns NOTES, drawn on the notebook. Lighthouse reports 4.36:1 on the original. After changing only the colour, all three scores become 100/100, while the keyboard barriers on A/B remain observed. agent-accessibility-tree passes in all six reports; the two applicable category audits pass (accessibility tree and layout stability), while the other five are not applicable. This result does not cover the Tab path.

Why 96 for B and 95 for C? The contrast failure has weight 7 out of 129 applicable points for A/C. B adds four passing ARIA audits with combined weight 28: (157 − 7) / 157 ≈ 95.54%, versus (129 − 7) / 129 ≈ 94.57%. The displayed score is rounded.

Control focusability (focusable-controls) is a manual check, excluded from the automated score.

Lighthouse 13.5.0 : agent-accessibility-tree.js

Method and sources

Measured environment

• Chrome 155.0.8059.39 · macOS arm64 · headless · 1280 × 900
• Node.js v24.19.0 · Playwright 1.63.0
• @playwright/mcp 0.0.68 · Playwright 1.59.0-alpha-1771104257000
• Lighthouse 13.5.0

Sources frozen at commit ae880251a7e8b40ceaed41e4eca2b7b82137d762.

Reproduce the measurements · measure.mjs · package-lock.json · measurements.json · Evidence-pack checksums

git clone https://github.com/edikkaweb/html-aria-agent-demo.git
cd html-aria-agent-demo/preuves/2026-10-08
npm ci
npm run measure

Node.js 22.19 or later and Google Chrome are required. New output goes to runs/DATE without replacing the published reports. See README for the keyboard test that exits with an error on A/B.

Raw reports

Original sources

• Variant A : readable report · JSON data · keyboard tests
• Variant B : readable report · JSON data · keyboard tests
• Variant C : readable report · JSON data · keyboard tests

Notebook colour adjusted

• Variant A : readable report · JSON data · keyboard tests
• Variant B : readable report · JSON data · keyboard tests
• Variant C : readable report · JSON data · keyboard tests

MCP

• Original sources : MCP A · MCP B · MCP C
• Cursor changed to default : MCP A · MCP B · MCP C

Limitations

One run per condition, Chrome on macOS, three simple pages. Measurements were run locally on frozen sources, then published here. No screen-reader or autonomous-agent trial, and no full conformance conclusion. Publishing these files is not a new trial against the live site.