← 回到实验档案

EXPERIMENT / 004 AI 协助 · 多轮修复

Meridian,精密的层次。

经过试玩反馈和多轮修复,本地模型能把交互作品做到什么程度?

大家怎么选

你的眼光,和大家一样吗?

正在读取票数…

Qwen 3.8 · 27B

—
—

Ornith 1.5 · 35B

—
—

无需登录 · 每个浏览器一票,可改选。先看看下面的作品。

匿名偏好投票,不是能力排行榜。
Qwen 3.8 · 27B本地模型
Ornith 1.5 · 35B本地模型
Astra 云端参考 · 不参与投票
Astra · reference

BEHIND THE BUILDS

做出来之后,差别在哪里?

最终作品都通过了已列出的基础检查。真正留下差异的,是这次把作品修到可展示的过程。

修复轮次产生代码修改的轮次
Qwen 3.8 · 27B2 轮

动画、指针与移动端布局

Ornith 1.5 · 35B8 轮

续写后,修复布局、运动与几何

已记录的修复耗时分钟 : 秒
Qwen 3.8 · 27B17:07
Ornith 1.5 · 35B33:37

含模型加载与工具等待;不含初始生成、续写及人工检查。

ASTRA / CLOUD REFERENCE

云端参考单独展示,使用 Codex;未记录可比耗时,不进入上方对比。

检查与来源基础检查 · 运行档案 +

以下是修复后版本的功能检查,不是区分模型的评分。

修复后版本的基础检查
检查Qwen 3.8 · 27BOrnith 1.5 · 35BAstra reference
JavaScript 语法✓已通过✓已通过✓已通过
展开与复原✓已通过✓已通过✓已通过
键盘旋转✓已通过✓已通过✓已通过
部件选择✓已通过✓已通过✓已通过
暂停与重置✓已通过✓已通过✓已通过
390px 布局:无横向滚动✓已通过✓已通过✓已通过
测试中无运行时错误✓已通过✓已通过✓已通过

OpenCode 1.18.25 + oMLX 0.7.0rc1。各轮反馈和修复记录留在档案中;页面上方展示最终作品,下方只展示初始任务。

Qwen 3.8 · 27B43.9 KB

Qwen 根据试玩反馈修复动画、指针、层级和手机布局。当前作品为模型修改的版本;先前的评审 CSS 分支仅作历史留档。最后一轮在所需修改保存并检查后结束。 耗时仅合计产生修改的修复轮次,含加载和工具等待,不包含人工检查及首轮。

下载作品源码 ↓

Ornith 1.5 · 35B52.8 KB

Ornith 根据浏览器反馈完成布局、部件对齐、运动状态和几何修复;后几轮包含评审给出的具体替换指令。服务连接失败单独记录。这是经过协助的多轮开发成果。 耗时仅合计产生修改的修复轮次,含加载和工具等待,不包含人工检查及首轮。

下载作品源码 ↓

Astra reference22 KB

使用不同 harness;首版与修订分别保留。

下载作品源码 ↓

THE ORIGINAL BRIEF

一份任务,同一个起点。

PROMPT.TXT

两份本地作品,从同一份初始任务开始。后续修复不同,当前展示的是修复后的结果。

初始任务 · 原文
下载 .txt ↓
Build a premium interactive single-page website for Meridian, a fictional mechanical watch. Deliver one complete index.html with inline CSS and vanilla JavaScript. No remote requests, external assets, fonts, CDNs, frameworks or 3D libraries. Build the three-dimensional assembly with CSS 3D transforms and inline SVG/HTML geometry.

Design direction: FUTURE PRECISION INSTRUMENT. Graphite or cool gray environment, cold silver metal, translucent glass, restrained ice-blue functional accents. A substantial, believable watch case, consistent light, clear depth and occlusion. The first screen should feel like a refined product presentation, not a control-panel demo. Exploration should reveal the precision mechanism. Use disciplined typography, whitespace and subtle motion, not neon glow or decorative gradient overload. You choose the composition and visual details. Explicitly label the mechanism as a simplified interactive illustration, not a real certified movement.

Required behavior:
1. Assembled watch with case, bezel, dial, generated hour markers and hour/minute/second hands showing real current time.
2. Direct drag rotation of the entire assembly. Touch and keyboard users must also be able to rotate it.
3. Smooth, reversible Explode / Reassemble controls. Separate case, dial, hands and individual movement parts so the structure becomes readable. Rotation must still work while exploded.
4. A visible or deliberately accessible mechanical movement: coiled mainspring barrel, balance wheel that oscillates back and forth, at least three distinct rotating gears. Gear speeds must be calculated from their teeth/size relationships, not identical arbitrary rotation.
5. Select individual components to read an accurate short name/function label. Selection must work on touch, not only hover. Selected components should be distinguishable.
6. Reset view and pause motion controls. Respect reduced motion or offer an immediately accessible pause; core controls must remain usable.
7. Usable at 1440x900 and 390x844: no page-level horizontal overflow, readable labels, keyboard-operable controls and visible focus.

Evaluation: real interactions, repeatable state/reset, no runtime errors, offline operation, keyboard/touch access, responsive layout, and source review. Visual taste will be judged separately by people. Do not invent benchmark scores or claim that tests were run.

Create index.html in the CURRENT EXPERIMENT DIRECTORY. Do not inspect parent directories or other projects. All requirements are in this prompt; there are no reference solutions to read. Use the available file tools to create the file. No commentary is necessary after the file is written.

版本与实验说明

本地模型使用同一份初始任务,随后分别补全与评审。原始输出、补全 Prompt 和评审改动均保留;配置与修复次数不同。Astra 使用 Codex,仅作为云端参考,不参与本地模型投票。

展示环境禁止联网,单次实验只说明这一次的结果。

继续挑选作品