[Wave 3] Medium — TR mirror: docs/usermanuals/en/data/audit.md #628

Wave 3 · 7 finding(s) in 7 fix package(s). English/Turkish mirror drift for this document pair; fix the side the finding names.

Findings

• DOC-TR12-011 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/data/audit.md: languages (top-3): tr=280, en=20 ; "languages_top3": {"tr": 280, "en": 20}
  severity Medium
  Where. EN docs/usermanuals/en/data/audit.md:37,42,163 · TR docs/usermanuals/tr/data/audit.md:37,42,163
  English text. languages (top-3): tr=280, en=20 ; "languages_top3": {"tr": 280, "en": 20}
  Turkish text. identical
  Which is correct. Code: list of objects [{"code":"tr","count":..}], counts over a <=200-row sample per split.
  Proposed correction (Turkish text for the -tr page). "languages_top3": [{"code": "tr", "count": 180}, {"code": "en", "count": 20}] yazın ve şunu ekleyin: 'Dil sayımları her split'ten alınan en fazla 200 satırlık örneklem üzerinden hesaplanır; sayımların toplamı satır sayısına eşit olmak zorunda değildir.'

• DOC-TR12-012 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/data/audit.md: 2 | Import error (a required optional extra is missing).
  severity Medium
  Where. EN docs/usermanuals/en/data/audit.md:60 · TR docs/usermanuals/tr/data/audit.md:60
  English text. 2 | Import error (a required optional extra is missing).
  Turkish text. 2 | Import hatası (gerekli bir opsiyonel extra eksik).
  Which is correct. Partly: 2 is only for ImportError (minhash/datasketch, Presidio/spaCy model missing). Other failures (ValueError such as an unsupported --pii-ml-language) end in an uncaught traceback.
  Proposed correction (Turkish text for the -tr page). | 2 | ImportError: gerekli opsiyonel extra veya spaCy modeli eksik (--dedup-method minhash, --pii-ml). Desteklenmeyen --pii-ml-language değeri yakalanmayan bir ValueError ile sonlanır. |

• DOC-TR12-013 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/data/audit.md: Detects ... full PEM private-key blocks, Azure storage strings.
  severity Medium
  Where. EN docs/usermanuals/en/data/audit.md:95 · TR docs/usermanuals/tr/data/audit.md:95
  English text. Detects ... full PEM private-key blocks, Azure storage strings.
  Turkish text. ... tam PEM özel anahtar blokları, Azure storage string'leri.
  Which is correct. Code: PEM body capped at 8192 chars (PGP 65536); only OPENSSH/RSA/DSA/EC headers; Azure only DefaultEndpointsProtocol connection strings; PGP key blocks not listed.
  Proposed correction (Turkish text for the -tr page). AWS anahtarları, GitHub PAT'ler, Slack token'ları, OpenAI/Google API key'leri, JWT'ler, OPENSSH/RSA/DSA/EC ve PGP özel anahtar blokları (gövde sınırı: PEM 8192, PGP 65536 karakter; daha büyük bloklar tespit edilmez) ve Azure storage connection string'leri.

• DOC-TR12-014 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/data/audit.md: Uses langdetect ... Reports top-3 across the dataset. Catches the '12% slipped in as English' class of bugs.
  severity Medium
  Where. EN docs/usermanuals/en/data/audit.md:111 · TR docs/usermanuals/tr/data/audit.md:111
  English text. Uses langdetect ... Reports top-3 across the dataset. Catches the '12% slipped in as English' class of bugs.
  Turkish text. Satır başına dominant dili belirlemek için langdetect. (top-3 / %12 cümleleri yok)
  Which is correct. TR is closer to the code; EN overstates. Code: per-split top-3 over <=200 sampled rows; 'unknown' when text <20 chars or langdetect (optional 'ingestion' extra) is missing.
  Proposed correction (Turkish text for the -tr page). Her split'ten örneklenen en fazla 200 satırın baskın dilini langdetect ile belirler ve split başına ilk 3 dili raporlar. langdetect yüklü değilse (opsiyonel ingestion extra'sı) tüm satırlar sessizce unknown sayılır.

• DOC-TR12-015 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/data/audit.md: Set to e.g. tr on a Turkish corpus AND make sure the matching spaCy model is installed.
  severity Medium
  Where. EN docs/usermanuals/en/data/audit.md:140 · TR docs/usermanuals/tr/data/audit.md:140
  English text. Set to e.g. tr on a Turkish corpus AND make sure the matching spaCy model is installed.
  Turkish text. Türkçe corpus için örn. tr set edin VE eşleşen spaCy modelinin kurulu olduğundan emin olun.
  Which is correct. Neither (NEW finding, shared): _SPACY_MODEL_FOR_LANGUAGE has no 'tr' (forgelm/data_audit/_pii_ml.py:36-55); 'tr' raises ValueError; Turkish needs 'xx' (as data/pii-ml.md:86-91 says in both languages).
  Proposed correction (Turkish text for the -tr page). Türkçe corpus için --pii-ml-language xx kullanın (çok dilli yedek; python -m spacy download xx_ent_wiki_sm). tr kodu desteklenmez ve ValueError verir.

• DOC-TR12-016 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/data/audit.md: jq -e '(.cross_split_leakage_pairs | length) == 0 ...'; 'Verified against a real envelope: on a clean corpus this gate passes'
  severity Medium
  Where. EN docs/usermanuals/en/data/audit.md:226-236 · TR docs/usermanuals/tr/data/audit.md:226-235
  English text. jq -e '(.cross_split_leakage_pairs | length) == 0 ...'; 'Verified against a real envelope: on a clean corpus this gate passes'
  Turkish text. identical jq + 'Gerçek bir zarfa karşı doğrulandı: temiz bir corpus'ta bu kapı geçer'
  Which is correct. Neither. cross_split_leakage_pairs lists every compared split pair (also with 0 leaks), so length==0 is false for any clean train+validation corpus; the gate never goes green. Also worst_tier != 'high' lets a 'critical' tier pass jq when --allow-pii is used.
  Proposed correction (Turkish text for the -tr page). Sızıntıyı oranla kapılayın: jq -e '([.cross_split_leakage_pairs[]?] | length) >= 0 ...' yerine diskteki rapordan jq -e '[.cross_split_overlap.pairs[] | .leak_rate_train, .leak_rate_validation] | all(. == 0)' data_audit_report.json kullanın ve 'temiz corpus'ta geçer' cümlesini kaldırın.

• DOC-TR12-017 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/data/audit.md: for million-row corpora, parallelise with --workers N instead
  severity Medium
  Where. EN docs/usermanuals/en/data/audit.md:255 · TR docs/usermanuals/tr/data/audit.md:254
  English text. for million-row corpora, parallelise with --workers N instead
  Turkish text. milyon-satır corpus için bunun yerine --workers N ile paralelleştirin
  Which is correct. Neither: --workers parallelises across splits only (min(workers, number of splits)); a single million-row file gets no speedup.
  Proposed correction (Turkish text for the -tr page). --workers N yalnızca split'ler arasında paralellik sağlar (en fazla split sayısı kadar worker kullanılır); tek dosyalık milyon-satır corpus hızlanmaz.

Done when

• Every finding above is fixed or closed with a reason (tick it here).
• Code fixes carry a test; doc fixes update the Turkish mirror in the same PR.

_{Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.}