Web → code → Web
We let an agent inspect a captured real website, write a reconstruction, then score the result against the original across visual fidelity, DOM/content structure, and interaction behavior. This site is the static viewer for the current realistic benchmark.
Best & worst realistic reconstructions
—
Checkpoint diff grid current wclone sample x 2 viewports x 3 steps · click any cell to compare
< 20%
20 – 45%
≥ 45%
realistic tasks
—
No-AJ runs
—
mean primary
—
categories
—
Task Browser
All realistic benchmark tasks, interaction scoring cases, submissions, and score breakdowns.
browse →Compare
Slider view over every captured/scored interaction step, with reference vs agent screenshots.
open →Diff View
Step through the visual mismatch overlay for reconstructed website states.
open →Data Index
Static JSON index containing task metadata, interactions, submissions, and score summaries.
open →How it works
-
Website interaction — each task packages a captured real
website plus replay states and private interaction/scoring cases under
wclone-realistic-benchmark/. - Code reconstruction — the agent uses the available website evidence to write a new HTML reconstruction rather than copying the server artifact verbatim.
- Scoring — the scorer compares reference and agent states across visual similarity, DOM/content fidelity, and explicit interaction outcomes. The viewer exposes the task, interaction cases, submissions, score breakdowns, and per-step visual comparisons.