Tidas Bilingual Transcreation

TIDAS Bilingual Transcreation

Scope

• Produce reviewed bilingual translations for local TIDAS rows after dataset bilingual extract only when a task explicitly requests post-import language completion.
• Do not run this skill for external dataset import readiness; import-ready rows are source-language rows.
• Use TIDAS/ILCD context, process or flow QA artifacts, source evidence, terminology, units, classification, geography, time, and technology context.
• Do not translate by mechanical term replacement. Write target-language text that is natural, professional, and faithful to the dataset meaning.
• Do not perform remote writes, publish, save-draft, or credential handling from this skill.

Canonical CLI Flow

Run the deterministic CLI steps around the AI review:

tiangong-lca dataset bilingual extract \
  --input /abs/path/rows/processes.jsonl \
  --type process \
  --out-dir /abs/path/translation/process

Codex then reads outputs/trans-units.jsonl plus the relevant evidence and writes a reviewed JSONL file:

{"unit_id":"...","row_index":0,"field_path":"/json_ordered/processDataSet/.../name","source_lang":"en","target_lang":"zh","source_text":"Electricity mix, high voltage","translated_text":"高压电力组合","basis":"Translated from process name, geography, and quantitative-reference context.","review_status":"agent_reviewed","reviewer":"codex"}

Apply and validate:

tiangong-lca dataset bilingual apply \
  --input /abs/path/rows/processes.jsonl \
  --translations /abs/path/translation/process/trans-reviewed.jsonl \
  --out /abs/path/rows/processes.translated.jsonl \
  --out-dir /abs/path/translation/process/apply

tiangong-lca dataset bilingual validate \
  --input /abs/path/rows/processes.translated.jsonl \
  --type process \
  --out-dir /abs/path/translation/process/validate

Repeat for --type flow when reference flows are part of the target dataset.

Translation Rules

• Preserve numeric values, units, CAS numbers, standards, UUID-like identifiers, geography codes, dates, and named methods unless the evidence explicitly supports a localized form.
• Translate the whole field in context. For example, avoid literal fragments such as 组件s, 层压件s, or 光伏 装置.
• Keep technical nouns precise: use terms such as 过程, 产品流, 基本流, 定量参考, 高压电力组合, or 光伏安装系统 only when they fit the actual field context.
• Keep source meaning narrower than style preference. Do not add process steps, environmental mechanisms, supplier claims, or assumptions not present in source evidence.
• If the English source is itself low quality, repair the target text only within supported meaning and record the limitation in basis.
• If the correct translation depends on missing evidence, output no translation for that unit and record a manual_review blocker outside the reviewed JSONL.

Required Inputs

Use the smallest set needed for the requested rows:

• outputs/trans-units.jsonl from dataset bilingual extract.
• Source rows or selected field snippets for field-level context.
• Source evidence, if available.
• Process/flow QA reports, if available.
• Flow governance decisions and process-flow reference rules, if the field names or comments mention linked flows.
• Project terminology or user-approved glossary, if provided.

Output Contract

Each reviewed translation row should include:

• unit_id
• row_index
• field_path
• source_lang
• target_lang
• source_text
• translated_text
• basis
• review_status, normally agent_reviewed
• reviewer, normally codex

The CLI writes translation-evidence.json during dataset bilingual apply; do not hand-edit that evidence file except to inspect it.

Stop Rules

Stop and report blockers instead of applying translations when:

• source evidence is missing for a high-impact field such as name, quantitative reference, geography, time, technology, or exchange meaning;
• translation would require inventing technical facts;
• a field contains mixed-language residue, placeholders, or schema-invalid text that needs content repair before translation;
• dataset bilingual validate reports blockers after apply.

The language-completion task is not done until dataset bilingual validate passes and translation-evidence.json exists for all reviewed bilingual fields. This is not an import or publish readiness requirement.