Mixed Locale Metadata — Guardrails and Fix Pattern

🧭 Quick Return to Map

You are in a sub-page of LanguageLocale.
To reorient, go back here:

LanguageLocale — localization, regional settings, and context adaptation

WFGY Global Fix Map — main Emergency Room, 300+ structured fixes

WFGY Problem Map 1.0 — 16 reproducible failure modes

Think of this page as a desk within a ward.
If you need the full triage and all prescriptions, return to the Emergency Room lobby.

A focused fix when language, script, or region metadata is inconsistent across files, snippets, headers, analyzers, or indexes. Use this page to normalize BCP-47 tags, align analyzers and encoders, and lock the schema so retrieval stays stable.

Open these first

Why this snippet and how it was chosen: Retrieval Traceability
Contract fields you must carry end to end: Data Contracts
Tokenizer picked the wrong rules: tokenizer_mismatch.md
Sources mix scripts in one turn: script_mixing.md
Full and half width digits or punctuation drift: digits_width_punctuation.md
Accents and folding collapse meaning: diacritics_and_folding.md
Locale choice flips between runs: locale_drift.md
CJK segmentation and wordbreak checks: cjk_segmentation_wordbreak.md
RTL and BiDi control safety: rtl_bidi_control.md
Transliteration pipelines and romanization: transliteration_and_romanization.md

When to use this page

Corpus claims lang=en in headers while HTML meta, PDF tags, or file paths imply non-English.
Index analyzer set to standard but crawler stamped lang=zh-Hant or lang=ja.
Mixed lang values inside one document after conversions or merges.
Query comes from a UI locale while the text is in another, leading to wrong analyzer or stopword set.
Re-ingest does not update stale lang/script/region fields and recall becomes unstable.

Acceptance targets

ΔS(question, retrieved) ≤ 0.45 on three paraphrases.
Coverage ≥ 0.70 to the intended section.
λ stays convergent across two seeds after metadata normalization.
E_resonance flat on long windows after analyzer alignment.

The core problem

Most stacks carry only lang and drop script and region, or they conflate UI locale with content locale. Encoders and analyzers then guess. That guess diverges from retrieval, chunking, and ranking, which raises ΔS and flips λ.

60-second fix

Detect
- Run lightweight LID on both raw source and post-parser text. Keep the probability.
- If p_top − p_second < 0.20, mark the snippet mixed and route to re-chunk.
Normalize
- Stamp BCP-47 tags as three fields: lang, script, region where known.
- Normalize Unicode to NFC for CJK and most Latin, NFKC only when you must collapse width.
Align analyzers and encoders
- Choose analyzer by (lang, script) rather than UI locale.
- Record analyzer_id, encoder_id, normal_form in the snippet.
Verify
- Re-rank with a deterministic reranker.
- Recompute ΔS and λ across 3 paraphrases, two seeds. Ship if targets pass.

Minimal contract you must carry

Add these keys to your snippet schema. See the full table in Data Contracts.

{
  "snippet_id": "doc:123#p4",
  "lang": "zh",
  "script": "Hant",
  "region": "TW",
  "normal_form": "NFC",
  "analyzer_id": "icu_zh_Hant",
  "encoder_id": "e5-multilingual",
  "lid": { "label": "zh-Hant", "p": 0.93, "method": "cld3|fasttext" },
  "source_meta": { "declared_lang": "en-US", "html_meta_lang": "zh-Hant", "pdf_lang": "zh" }
}

Rules

Prefer content-derived LID over HTTP headers or file paths.
If declared_lang disagrees with LID, trust LID and record the disagreement.
Never mix snippets with different (lang, script) into the same chunk group.

Quick diagnosis → fix

Answers oscillate per run although facts exist → The analyzer follows UI locale. Lock analyzer by snippet fields. See Retrieval Traceability.
High similarity yet wrong meaning → Encoded with a multilingual model but index analyzer used wrong token rules. Rebuild with aligned analyzer. See embedding-vs-semantic.md.
Mixed half width, punctuation, or digits break matching → Normalize to NFC or NFKC per policy and re-index. See digits_width_punctuation.md.
Script switches mid-paragraph → Split at the switch and stamp each part. See script_mixing.md.
CJK recall dips after parser change → Check wordbreak rules and ICU configs. See cjk_segmentation_wordbreak.md.

Copy-paste prompt for your LLM step

I uploaded TXT OS and WFGY Problem Map.

My issue looks like mixed locale metadata:
- declared vs detected lang differ,
- analyzer_id does not match (lang, script),
- ΔS(question, retrieved) = {value}, λ = {state}

Tell me:
1) which layer is failing and why,
2) the exact WFGY page to open,
3) the minimal steps to align BCP-47, analyzer, and encoder,
4) a reproducible test that verifies ΔS ≤ 0.45 and λ convergent.
Use BBMC/BBCR/BBAM when relevant.

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

Explore More

Layer	Page	What it’s for
Proof	WFGY Recognition Map	External citations, integrations, and ecosystem proof
Engine	WFGY 1.0	Original PDF based tension engine
Engine	WFGY 2.0	Production tension kernel and math engine for RAG and agents
Engine	WFGY 3.0	TXT based Singularity tension engine, 131 S class set
Map	Problem Map 1.0	Flagship 16 problem RAG failure checklist and fix map
Map	Problem Map 2.0	RAG focused recovery pipeline
Map	Problem Map 3.0	Global Debug Card, image as a debug protocol layer
Map	Semantic Clinic	Symptom to family to exact fix
Map	Grandma’s Clinic	Plain language stories mapped to Problem Map 1.0
Onboarding	Starter Village	Guided tour for newcomers
App	TXT OS	TXT semantic OS, fast boot
App	Blah Blah Blah	Abstract and paradox Q and A built on TXT OS
App	Blur Blur Blur	Text to image with semantic control
App	Blow Blow Blow	Reasoning game engine and memory demo

If this repository helped, starring it improves discovery so more builders can find the docs and tools.

9.7 KiB Raw Blame History Unescape Escape