Explore an isolated website
The Agent receives a live browser surface and a task contract. It does not receive the reference implementation, source tree, database, or private fixtures.
Routes · interactions · server state
Evaluation protocol
Reconstruct from interaction, then explain every difference
The benchmark measures whether an Agent can infer and rebuild a stateful website from a bounded offline server. Dataset acceptance proves the input is trustworthy; only a later Agent run produces a benchmark score.
Infer and implement
The Agent navigates, tests hypotheses, writes code, and produces a separately runnable clone under the same isolation rules.
Trace · files · runtime · usage
Compare the new clone
Automated and multimodal judges compare observable behavior and appearance against the standard offline reference.
Screenshots · replay · terminal state
Scoring evidence
Five dimensions, each backed by inspectable artifacts
Visual fidelity
Matched screenshots across routes, states, and viewports; differences can be inspected with split, overlay, blink, and heatmaps.
Interaction fidelity
Clicks, typing, menus, option dependencies, mutations, validation, and before/after states behave like the reference.
Journey completion
Success, failure, and recovery journeys reach the correct durable terminal states instead of only looking plausible.
Robustness
Hidden seeds, invalid inputs, retries, persistence, reset behavior, and network isolation test behavior beyond the happy path.
Efficiency
Elapsed time, tool use, tokens, retries, and runtime resources contextualize the quality of the reconstruction.
Before experiments
Dataset calibration
Original website evidence is compared with the standard offline reference to certify that the benchmark input is representative, runnable, stateful, and isolated.
Original evidence↔Offline reference
Never an Agent scoreDuring experiments
Agent evaluation
The accepted offline reference is compared with a newly generated clone. This evidence produces per-dimension results and the official Agent score.
Offline reference↔Generated clone
Populates the score matrixWhat remains blank today
Agent identities
No experiment has declared a model, provider, harness, or reasoning configuration.
Generated clones
No Agent-produced runnable server has been submitted for comparison.
Official scores
Visual, interaction, journey, robustness, efficiency, and total values are intentionally empty.
Run evidence
Agent traces, generated-clone screenshots, failure records, and usage metrics will appear only with real reports.