Define quality before translating

A fluent sentence can still be commercially wrong. Professional localization must preserve meaning, facts, terminology, tone, calls to action and technical structure. Create an acceptance standard before the first bulk job.

Recommended Professional configuration

• Quality preset: Professional.
• Creativity/variation: 0.1 baseline.
• Independent AI quality-review pass: enabled.
• Ambiguous or risky content: keep in review.
• New target status: Draft until the workflow is proven.
• Qualified model score: at least 90/100.

Build the glossary

1. List product names, company names, feature names and registered marks.
2. Specify terms that must remain unchanged.
3. Add approved source-to-target terminology for recurring concepts.
4. Define capitalization and punctuation rules.
5. Add formal/informal address and regional language requirements.
6. Resolve contradictory entries before saving.

Keep the glossary precise. Do not add whole marketing paragraphs as terms. If a source term has multiple valid translations, explain the context or leave it for review instead of forcing one meaning everywhere.

Write useful project instructions

State the target market, audience, tone, reading level, units, currency display, prohibited claims and words that must not be translated. Avoid vague requests such as “make it better.” Instructions should not conflict with the source, glossary or protected-token rules.

Structural safeguards

Translator protects placeholders, shortcodes, HTML tags, URLs, email addresses, IDs, code tokens and structured builder syntax. The qualification and result validators compare protected values and reject structurally unsafe output. A rejected result is safer than saving fluent text that breaks the page.

Independent review

The review pass evaluates the translation against the source, not merely as standalone prose. It checks missing meaning, added claims, wrong numbers, terminology, language, placeholders and structural validity. Review does not remove the need for a human decision on legal, medical, financial, regulated or brand-sensitive content.

Human review procedure

1. Open the result card in Jobs & Logs.
2. Read the completion message and technical details.
3. Preview the target beside the source.
4. Review headings, calls to action, numbers, units, links and product claims.
5. Check glossary terms with a site search.
6. Edit the target-language item directly.
7. Test the front end and publish only after approval.

Common quality failures

• Wrong language: verify the row’s original language, selected target and multilingual assignment.
• Brand term translated: add a protected glossary rule and retry the affected target.
• Placeholders changed: use a stronger qualified model and inspect conflicting instructions.
• Invented claims: lower variation, require review and keep unsupported content in Draft.
• Inconsistent terminology: approve glossary and memory entries, then update only outdated segments.
• Awkward button text: edit for the target UI while preserving intent and link behavior.

Approval checklist

• No meaning, number or limitation is missing.
• No new claim was introduced.
• Brand and glossary terms are consistent.
• Protected tokens and links are intact.
• The text fits the actual interface.
• SEO text describes the target page accurately.
• High-risk content received qualified human review.