← ClaudeAtlas

parity-probelisted

Compare a bilingual documentation pair (README.md/README.de.md, SECURITY.md/SECURITY.de.md) for structural parity — heading skeleton, per-section bullet counts, tagged code blocks, link targets, and how far the translation lags in git. Never compares prose, which is supposed to differ. Deterministic; run it, do not reason about it.
malkreide/mcp-continuous-auditor · ★ 0 · AI & Automation · score 71
Install: claude install-skill malkreide/mcp-continuous-auditor
# Parity probe Does the translated documentation still say the same thing? ```bash python scripts/parity_probe.py --target <path> python scripts/parity_probe.py --target <path> --lang fr --format json python scripts/parity_probe.py --target <path> --pair README.md:README.de.md ``` Exit `0` parallel, `2` **findings**, `3` no translated documents (NOT MEASURED), `4` the checkout moved during the run, `127` the harness could not run. ## Why The portfolio is bilingual. The English side is where features get written up, so it moves first; the German side is what a Swiss reader opens first. Both files render, so a section that exists in one and not the other is invisible without reading both and counting. ## What is compared, and what is not | Compared | Not compared | |---|---| | the sequence of heading **levels** | heading **text** — `Overview` / `Übersicht` is a correct translation | | top-level list items per section | nested items — a translator may split a sub-point | | fenced blocks that **declare a language** | untagged fences — a directory tree is prose in a monospace font | | commands inside those blocks | comments inside them, whole-line and trailing | | link targets | the cross-language link itself, in both directions | `TRANSLATION_LAG` counts commits that touched the base file after the last commit that touched the translation. It is the only check that fires when every structural check is green — which is what a half-translated paragraph looks like. Updating