Why

Refactoring HTML/CSS under UI preservation (extracting components, migrating design systems, upgrading component libraries, removing dead CSS) changes the source while intending to preserve what the user sees. Pixel diffing is flaky and non-structural; CSS diffing flags the very changes refactoring is meant to make; manual QA does not scale. sosie exists to give a trustworthy "equivalent / not equivalent" verdict.

Design principle: trustworthiness first

The value of sosie is the strength of its "equivalent" verdict. The design is built around an asymmetry:

• False negatives (missed visual regressions) are fatal. A single one reaching production destroys trust.
• False positives (spurious diffs) are tolerable. The user inspects, sees it is harmless, and moves on.

The property whitelist is the explicit trust boundary: sosie's verdict is exactly as strong as the whitelist is complete.

Build

The project uses a local opam switch (_opam/, gitignored). Use opam exec -- to enter the environment:

opam exec -- dune build
opam exec -- dune runtest
opam exec -- dune exec sosie -- --version

Usage

sosie is a CLI with several subcommands:

sosie capture           # Capture a DOM snapshot from a URL
sosie compare           # Compare two DOM snapshots
sosie capture-all       # Batch capture: routes x viewports x color schemes
sosie audit-whitelist   # Detect blind spots in the CSS property whitelist
sosie show-config       # Visualize normalization config overlays on a page

Run sosie <command> --help for the full options of each subcommand.

Documentation

• Design: sosie-design.md
• Roadmap: sosie-roadmap.md
• Formal methods survey: formal-methods-survey.md
• Contributor conventions: CLAUDE.md

AI disclosure

In the interest of transparency, and in keeping with the transparency principles of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689):

• This repository was developed with substantial assistance from a generative AI system — Anthropic's Claude, used through the Claude Code agentic coding tool. AI assistance covered source code, tests, documentation, and design discussion. All AI-produced content was reviewed and is maintained under human direction; the maintainer is responsible for the contents of this repository.
• sosie itself is not an AI system. It contains no machine-learning models and performs no AI inference at runtime. Its output is a deterministic structural comparison of DOM snapshots. Consequently, sosie does not fall within the scope of the AI Act's obligations for AI systems, and it produces no "AI-generated content" within the meaning of Article 50.
• This notice is a voluntary provenance disclosure. It is not a claim of formal certification or conformity assessment under the AI Act, which governs AI systems placed on the EU market rather than the use of AI as a development aid.

License

Released under the MIT License. © 2026 Cuihtlauac Alvarado.