Clone review — before / after
Side-by-side captures for the deterministic website cloner. Each image is one site: SOURCE (left) vs CLONE (right), with a desktop row (1280px, shown at 50%) on top and a mobile row (375px, 100%) below.
Results
| Tier | Gates 0–6 | Avg score | Detail |
|---|---|---|---|
| easy (24 personal/portfolio/content sites) | 24 / 24 | 99.7 | easy/RESULTS.md |
| medium (23 SaaS / ecommerce marketing pages) | 16 / 23 | 98.9 | medium/RESULTS.md |
| hard (32 dense editorial / docs / marketing / ecommerce) | 23 / 32 | 98.8 | hard/RESULTS.md |
| stage2 (25 popup / video / animation pages) | 22 / 25 | 99.1 | stage2/RESULTS.md |
| sites (5 multi-page sites, Stage 3) | 47 / 49 routes | — | sites/RESULTS.md |
Stage 2 ("capture the right state") adds two gates beyond 0–6 — pollution
(degenerate / bot-wall / un-dismissed-modal capture) and perceptual (tier-
thresholded screenshot diff) — plus capture-time overlay/popup dismissal,
animation settling, and dynamic-media first-frame, and a single-load+resize
capture that keeps content identical across viewports. 19/25 also clear the
stricter stage-2 bar (gates 0–6 + pollution + perceptual); browse stage2/.
Stage 3 ("multi-page / whole-site") clones a whole site from one entry URL into
one Next.js app: it crawls routes, collapses CMS collections to one representative
(jamstack 552→11, 11ty 1085→12 reproduced routes), hoists shared header/footer chrome
into the layout, rewrites internal links to the clone routes, and adds site-level
link-integrity + site-determinism gates. 47/49 routes pass gates 0–6, all 5
sites green on links + determinism; browse sites/.
Gates 0–6 = build, capture, asset/font, DOM, computed-style, layout, determinism.
Every non-passing medium site still scores 95–99; the misses are documented
limitations (see below). Screenshots are a human sanity check on top of the
quantitative gates — browse them in easy/ and medium/.
A few highlights
Pixel-perfect static reconstruction:
High-fidelity medium reconstruction:
Known limitations (visible in some medium shots)
A static, deterministic clone reconstructs the captured static state. Content driven by JS / time / scroll, or embedded from a third party, is non- deterministic frame-to-frame and cannot be reproduced — the static scaffolding (text, layout, materialized images) around it is still correct:
- Rotating hero text — shopify, notion (a JS timer swaps the word).
- Scroll-scrubbed / autoplay hero video — mercury (
<video>frame depends on scroll position; the fallback image is reconstructed, behind the video). <iframe>embeds — loom (dropped to keep the clone self-contained).- Fancy multi-line headings (per-word
inline-block+ spacers) stack a few px taller — linear, notion, deel.
Current architecture overview: ../README.md.
Generated app shape
The compiler emits a Next.js App Router + TypeScript project per site:
src/app/page.tsx (JSX reconstructed from the DOM), src/app/ditto.css
(clone-specific CSS that cannot be represented cleanly as utilities), and
layout.tsx. Larger clones additionally carry public/assets/ (materialized
images/fonts/icons/manifests), generated documentation, and optional
src/app/ditto/ runtime helpers. Regenerate any site with
cd ../compiler && npm run bench -- --tier=medium.