Scoring
Automated 60, visual 20, code 10, verification 5 and efficiency 5.
Repeated multi-brief evaluation
0 of 24 outputs qualified.
Round 1 compares raw quality across 24 independently generated websites. The corrected gate reanalysis is final and does not change the qualification outcome.
Category leaders
Different questions have different leaders.
Leader labels describe the frozen raw evidence. They are not a final qualification verdict.
Model results · 4 models
Raw performance, clearly separated from gates
Cards are ordered by aggregate raw score. Each aggregate equally weights Brand, Ops and Launch.
Scores by brief
Brand, operations and launch
Each cell is the median of two repetitions. Raw score leads; the recorded official score remains secondary.
Open detailed evidence tables
Component comparison
Where model profiles differ
Run-level detail
All 24 authoritative runs
Failed gates below are the original recorded outputs—not corrected qualification decisions.
Visual and code scores were produced by isolated, identity-blind AI reviewer agents. Targeted AI adjudication occurred only where the frozen disagreement rules triggered.
Methodology and limits
What these results mean
Repetition
Two runs per model and brief. Cell scores use medians; suite scores equally weight the three briefs.
Qualification
The original evaluator capped gate-failing runs at 49. The corrected reanalysis still leaves every output not qualified.
Limits
This is a small, AI-reviewed benchmark. Captured inference cost is not total operating cost.