Bez项目：利用AI从Web规范和测试自动生成浏览器引擎

Bez是一个通过人工智能模型从Web规范和测试自动生成浏览器引擎代码的项目1。该项目采用AI模型候选、引擎运行验证和跨浏览器对比的循环流程，从规范文本生成Web引擎代码，目标是降低构建Web引擎的成本1。

截至2026年9月25日，Bez已成功生成8条CSS 2.1布局规则，这些规则已通过227个recipe测试和11个WPT正常流页面测试验证1。项目采用多个数据源进行验证：WPT投票表格覆盖2,162,676个关键项（占总数的94.8%），同时通过对比Chromium、Firefox和WebKit三个浏览器的测试，实现了699/705的浏览器对比一致率（99.2%）1。

在实施过程中，Bez的验证流程还发现了Firefox浏览器的真实兼容性问题：Firefox在像素舍入时使用1/60的精度，而Blink和WebKit则采用1/64的精度1。根据项目评估，约55至60%的引擎相关兼容性条目拥有可用的自动化预言机和可生成的规范文本1。

Bez is an artificial intelligence-driven project that automatically generates browser engine code from Web specifications and test suites 1. The system employs a cyclic approach involving AI model candidates, engine execution verification, and comparative analysis across three major browsers—Chromium, Firefox, and WebKit—to validate generated code against both proprietary browser tests and the Web Platform Tests (WPT) collection 1.

As of September 25, 2026, the project has successfully generated eight CSS 2.1 layout rules and demonstrated substantial compatibility consistency across browsers 1. The WPT voting tables achieve 94.8% coverage with 2,162,676 key items, while cross-browser agreement rates reach 99.2% with 699 out of 705 test cases aligned 1. Generated rules have passed 227 recipe tests and 11 WPT normal flow page tests 1. The initiative aims to reduce the cost and complexity of building Web rendering engines by automating code generation from formal specifications.

The project has uncovered a real compatibility bug in Firefox, which uses 1/60 pixel rounding for layout calculations while Blink and WebKit employ 1/64 pixel rounding 1. Current analysis indicates that approximately 55 to 60 percent of engine-related compatibility entries possess usable automated test oracles and specification text suitable for code generation 1.