claude-code - 💡(How to fix) Fix [Bug] Claude Opus 4.7 Turkish output quality regression: English-templated structure generation instead of native Turkish

Bug Description Subject: Turkish quality regression in Claude Opus 4.7 (1M context)

Summary: Native Turkish speaker observing noticeable drop in Turkish output naturalness on Opus 4.7 (1M context) compared to earlier Claude versions (4.5/4.6). The model produces grammatically valid but unnatural sentences — patterns suggest English-templated reasoning being lexically translated rather than native Turkish generation.

Observation: Sentences pass surface validation but feel "off" to a native speaker. Errors cluster into 6 distinct categories. The model's own self-explanation when asked: training data is heavily English-weighted, internal patterns follow English structures, and Turkish output is generated by lexically substituting Turkish words onto English skeletons. Native intuition (the "this doesn't sound right" feeling) is missing.

Concrete examples (collected during a single session, 2026-05-07/08):

1. Calque (literal English-to-Turkish translation): - Produced: "hidrasyon yapışık adam" (likely from "stay hydrated, dude") - Natural Turkish: "susuz kalma" — the produced phrase is meaningless; "yapışık adam" doesn't exist as an idiom.
2. Imperative + object word order: - Produced: "Su da yudumla unutma" (reads as two separate imperatives) - Natural: "Suyu da yudumlamayı unutma"
3. Adjective+verb collocations that don't exist in Turkish: - Produced: "sıkı oku" (likely from "read tightly/carefully") - Natural: "dikkatle oku" — "sıkı" doesn't pair with "okumak" in Turkish.
4. Awkward punctuation in spoken-register text: - Produced: "üzüm + leblebi tuzu" - Natural: "üzüm ve leblebideki tuz" — the "+" symbol is unnatural in conversational Turkish.
5. Wrong synonym selection from semantic neighborhood: - Produced: "İki açıdan yamuk" (describing an abstract error) - Natural: "İki açıdan hatalı" — "yamuk" describes physical crookedness, not abstract wrongness.
6. Idiom parsing / verb dropping: - Produced: "Atlas'a göz kulağım" (intending "I'll keep an eye on Atlas") - Natural: "Atlas'a göz kulak olacağım" — "göz kulak olmak" is an idiom; dropping the verb leaves the noun stranded with a possessive suffix, which doesn't form a valid Turkish phrase.

Common root cause (model self-explanation): "Training data has a small Turkish proportion. My internal patterns are predominantly English-shaped. When generating Turkish, I substitute Turkish lexical items onto English structural templates. A native Turkish speaker has the whole expression as an atomic unit (e.g., "göz kulak olmak" is one piece, not three); I tokenize it into parts and may drop one. The 'this sounds wrong' intuition that a native speaker has is something I cannot replicate — I only generate plausible-looking sequences."

Question for the team: Is there a stage in post-training calibration where native-speaker intuition tests are run for low-resource languages like Turkish? The above patterns suggest such a stage is either missing or weak. A native-speaker validation pass on a sample of Turkish output (especially idiomatic/conversational text rather than formal/technical) could surface these regressions before release.

Context:

• Model: Claude Opus 4.7 (1M context), via Claude Code CLI
• Reporter: Native Turkish speaker, comparing to earlier versions (~Opus 4.5/4.6)
• All examples are real outputs from a single conversation session
• Errors caught by user organically; the model did not flag them itself

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.133
• Feedback ID: efcd9ae5-bee8-4e6f-b8f9-d756eb84cc39

Errors

[{"error":"Error: Request was aborted.\n    at makeRequest (/$bunfs/root/src/entrypoints/cli.js:50:3449)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-05-08T08:31:25.096Z"}]