Bez - a generated web engine#
Generate a web rendering engine from specs and tests.
Why#
Building a web engine by hand costs hundreds of engineers and many years, so only a few companies can have one, and only they decide how the web works. Bez generates the engine instead: from the specifications, with the three shipping browsers (checked against each other) and WPT as the tests. Once the pipeline exists, each extra engine costs very little.
Goals#
- Complete, or tree-shaken. The default build is a complete engine. For uses
other than a web browser, a build can instead contain only the web features its
content uses: analyse a site or an app, and every feature it never touches is left out
of the binary, not just switched off (
docs/content-scoped-engine.md). - Fast on any device. A smaller engine does less work and needs less memory.
- Easy to embed. A small engine with a clear API, for apps and devices that today ship all of Chromium or go without.
- Fastest to update. When a spec or test changes, the affected code is regenerated and re-verified, not rewritten by hand.
How it works#
flowchart LR
spec["spec text"] --> model["model writes<br/>many candidates"]
model --> engine["candidate runs<br/>inside the engine"]
browsers["Chromium, Firefox,<br/>WebKit (cached)"] --> check{"same boxes,<br/>same places?"}
engine --> check
check -- no --> model
check -- yes --> code["committed as<br/>ordinary Rust"]
docs/generation-loop.md— how one rule goes from spec text to committed code.docs/geometry-verification.md— how the engine's layout is checked against three browsers.docs/generated-functions.md— the layout rules generated so far and what each batch showed.roadmap.md— the plan, the decisions and the open questions.
Status#
Computed against browser-compat-data 8.0.4 (17259 leaf keys, 74 rules in docs/feature-map.json) on 2026-09-25.
| area | generated | hand-written | linked | oracle only | unreached |
|---|---|---|---|---|---|
| overall | 0.6% | 0.3% | 0.5% | 5.7% | 93.0% |
| css | 2.5% | 1.1% | 2.0% | 0.0% | 94.4% |
| html | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
| api | 0.0% | 0.0% | 0.0% | 9.9% | 90.1% |
| javascript | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
| svg | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
| webassembly | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
| http | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
| mathml | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
See docs/dashboard.md for what these statuses mean and how this table is kept in sync.
DOM, style, box tree and fragment tree are hand-written in crates/dom and
crates/layout and pass their browser checks
(roadmap.md → "Phase 1 — Engine bootstrap").
Nine CSS 2.1 layout rules live in
crates/layout/src/generated; eight were written by
models and admitted against the three-browser vote, and block height keeps its
hand-written rule because no model candidate beat it. Together they pass all 227 recipe
cases and all 11 usable WPT normal-flow pages. Which rules the engine actually calls:
roadmap.md → "Which resolver each generated function runs".
What would prove the thesis:
roadmap.md → "Proof of concept".
Findings so far#
- Three browsers mostly agree, and the third names the odd one out. 699 of 705
browser-pair comparisons agreed across 235 documents (commit
3ca7323); all six disagreements were twelve-deep percentage nesting, and majority voting named Firefox the outlier every time.docs/firefox-app-units.md. - That Firefox difference is a real web-compat bug. Gecko rounds lengths to 1/60 px
where Blink and WebKit use 1/64 px, which makes a flex item wrap only in Firefox or
offsetWidthdiffer by a pixel. Reproductions match Mozilla-diagnosed breakage on Slack, Google Store and Samsung, and Mozilla bug 1719314.experiments/firefox-compat/. - The WPT vote table agrees across most of the suite. A stable three-engine majority
covers 2,162,676 of 2,282,301 test/subtest keys (94.8%).
docs/wpt-votes.md. - Script-observable behaviour agrees too. Trace probes under a fake-media profile agree on 15 of 18 engine pairs; the three disagreements are a real platform difference. docs/dump-format.md → "Trace protocol".
- Conformance suites are usable oracles. Counting a test usable when two of three
engines agree: WPT canvas 82.6% (92.7% excluding tentative), Khronos WebGL with dEQP
99.7%, WPT Web Audio 74.4% (85.6% excluding tentative); the WebGPU CTS about 85% for
validation and about 50% for numeric execution.
docs/conformance-oracles.md,docs/webgpu-cts.md. - Logic programming earns a narrow place. Margin collapsing written as Datalog agreed
with the browsers on 1195 of 1195 offsets and caught a seeded bug the geometry check
missed in all 227 cases.
docs/logic-layer.md. - The economics hold for a subset of the platform. About 55–60% of engine-relevant
compat entries have a usable automated oracle and generatable spec prose; roughly 8–18%
have none.
docs/platform-economics.md,docs/oracle-coverage.md. - An engine scoped to one site's content is designed, and its content analyser has
landed.
docs/content-scoped-engine.md.
Open questions#
What building each rule costs by hand, how wide a feature's generated part should be,
how much of WPT is reachable without JavaScript, and where IDL-generated code ends:
roadmap.md → "Open questions". Area-by-area status:
docs/platform-areas.md.
Sources#
- browser-specs, every web spec in one repo: https://github.com/w3c/browser-specs
- web-features, the web platform grouped into features: https://github.com/web-platform-dx/web-features/
- browser-compat-data, every piece of the web surface: https://github.com/mdn/browser-compat-data
- WPT, the web platform tests: https://web-platform-tests.org/
- wpt-gen, WPT tests generated from specs: https://github.com/GoogleChromeLabs/wpt-gen
- webref, machine-readable terms from web specs: https://github.com/w3c/webref