vlmkit · demo sites
Sites built by agents, and how each one was judged
Each demo site was built from a one-page brief by an agent using vlmkit's gates and its own eyes, and the landing page was reviewed the same way. Next to each page is its judgment log: every gate run with its whole output, every screenshot that was looked at with what was seen in it, and whether each defect was found by a gate or only by looking.
- 7logged pages
- 645gate runs
- 1342screens looked at
- 127defects found by eye
- 43defects found by a gate
vlmkit landing page
product landing page (review) · English
- rounds
- 4
- gate runs
- 119
- screens seen
- 624
- found by eye
- 31
- found by a gate
- 3
- looks
- 108
Klondike solitaire
game — drag and drop and animation · English
Judged before the judgment log existed. Its README's table records which defects the gates, the tests and screenshots found — without the pictures.
Pantry Club checkout
multi-step form / checkout · English
- rounds
- 5
- gate runs
- 72
- screens seen
- 84
- found by eye
- 12
- found by a gate
- 4
- looks
- 47
Keel Analytics — Overview
SaaS analytics dashboard (app shell) · English
- rounds
- 5
- gate runs
- 56
- screens seen
- 64
- found by eye
- 11
- found by a gate
- 5
- looks
- 44
Quarry documentation
documentation site · English
- rounds
- 5
- gate runs
- 163
- screens seen
- 235
- found by eye
- 21
- found by a gate
- 8
- looks
- 63
Sprintboard
interactive app — kanban board with drag and drop · English
- rounds
- 5
- gate runs
- 89
- screens seen
- 68
- found by eye
- 13
- found by a gate
- 8
- looks
- 59
ほとり — 海辺の町の古本屋が、夜だけ店を開ける理由
editorial long-form article · Japanese
- rounds
- 6
- gate runs
- 71
- screens seen
- 150
- found by eye
- 26
- found by a gate
- 11
- looks
- 43
汐見窯 マグカップ 青磁
e-commerce product page · Japanese
- rounds
- 5
- gate runs
- 75
- screens seen
- 117
- found by eye
- 13
- found by a gate
- 4
- looks
- 60
How the logs were kept
Every log was written by one small tool as the work happened, under one protocol:
- a gate runs through it, so its exit code and its whole output are kept rather than a summary of them;
- a screenshot is taken through it, one screen per image, and a look is recorded against the image it describes; each finished round keeps the pictures of one full-page walk per width, so you can open them and disagree;
- a defect must name the look or the gate run that found it, and a fix must be checked by a later look or run before the log can be marked done.