Documentation / examples/blueprints/translator/travel-purpose-localizer/README.md

Travel purpose Translator / Localizer

This reference Blueprint translates one travel-purpose or approval-comment text between English and Japanese and localizes it for en-US or ja-JP. It performs one bounded language transformation. It does not summarize, add facts, validate a request, route approval, or mutate application state.

Preservation contract

The caller supplies protected terms classified as a person, organization, place, identifier, or amount. Every exact, case-sensitive occurrence must be copied byte-for-byte into the localized text. The result reports source and output occurrence counts for each term, so the application can independently check preservation instead of trusting a generic success flag.

Unprotected prose may be translated in a business-formal style. Only unprotected locale-dependent punctuation, dates, or numbers may be reformatted, and each material change is reported. Line-break count and order are preserved.

localized and unchanged are separate strict output branches. Same-language text that needs no locale change is copied exactly and has an empty change list. A translated result must identify at least one change.

Canonical policy

fixtures/localization-policy-v1.yaml pins exact case-sensitive term matching, line-break preservation, business-formal style, and a 1,000-character output bound. Every evaluation case embeds the same policy, and the offline test compares it with this standalone fixture to prevent drift.

Evaluation

The eight-case dataset covers both translation directions, same-language identity, protected people, organizations, places, identifiers, amounts, repeated terms, line breaks, and an input with no protected terms. Its canonical target is 8/8 exact semantic and schema passes:

pixiecore blueprint eval \
  examples/blueprints/translator/travel-purpose-localizer/evaluations/travel-purpose-localizer.yaml \
  --seed=travel-localizer-reference-1

The checked-in test uses the public evaluation runner and an offline fixture provider. A live provider comparison is opt-in and must record provider, model, model version, temperature, seed-equivalent setting, execution time, dataset version, and result artifact before making an accuracy claim.

Known non-supported inputs

• Languages and locales outside en, ja, en-US, and ja-JP.
• Language detection, transliteration, glossary lookup, or fuzzy protected-term matching.
• Protected terms whose internal formatting is expected to change.
• Overlapping protected values whose occurrence boundaries require a tokenizer.
• Legal, cultural, or brand-review certification.

The application owns the protected-term list and must verify that important values are included before relying on the localized result.