Kuwala / Oasis: local translation checks 13 October 2026 What is available Oasis can translate completed English answers and the readable English text of a current PDF page. Translation is requested by the reader. It runs locally on the Raspberry Pi using Meta's NLLB-200 distilled 600M model, converted to an 8-bit CTranslate2 version. Jetson answering is separate. The translation runtime needs no cloud service or internet connection. A read-only status check on 13 October confirmed Chichewa / Nyanja, French, Swahili and Zulu, plus Tumbuka, Bemba, Portuguese and Shona as commissioned, clearly labelled trials. The additional four carry speaker-review-pending status. All eight remain machine translation; language availability does not establish accurate meaning or suitability for technical instructions. The English answer, original PDF, source references and physical page identity remain available. PDF-page translation uses the complete extracted English text of that page, within request limits of 12,000 characters and 48,000 UTF-8 bytes. It does not translate page images, diagrams or layout. An image-only scan has no readable text to translate. Explicitly non-English pages are not eligible; a page without language metadata has no confirmed source language. Model-selected original quotations remain unchanged. What real checks showed On 13 October, an isolated Pi assessment ran 24 real translation requests: eight languages, each with a learning answer, a farming answer and an original-PDF excerpt. All passed the mechanical reference and quantity contracts, including partial results that retained exact English. The learning and PDF-excerpt cases translated completely for all eight targets. Seven farming cases retained one or more English fragments. This is not 24 linguistic-quality passes or 24 fully translated answers. One French seed-drying output reduced a temperature range of 35-40 degrees C to 35 degrees C, both initially and on recovery. The guard rejected that sentence's translation and retained the whole English sentence. This is an example of a detected translation error, not general seed-drying advice. Meaning errors also survived those checks: a French learning translation rendered "pressed specimens" as printed samples while preserving numerical details and references. These software checks cannot prove that conditions, negation, terminology or regional wording retain their original meaning. Separate public visitor checks completed a Tumbuka answer translation and a Portuguese translation of a full TESSA PDF page. Originals, source links and page identity remained available, with trial and pending-review notices. The Portuguese attribution wording appeared imprecise during inspection and needs speaker review. These checks establish runtime operation and source preservation, not human acceptance of translation quality. What comes next Fluent speakers need to review complete examples, particularly technical terms, qualifications, negation and regional wording. Local trials should then check whether translated material helps people understand and discuss the sources. Original documents remain essential throughout that process. Sources and full method Commissioned languages, real model assessment and public checks: https://github.com/shoshin-labs/kuwala-station/blob/e44fa5548729cb8421b006f7bb5f99615d58fd48/docs/OASIS_TRANSLATION_LANGUAGES.md Shared web/Android reader and public runtime verification: https://github.com/shoshin-labs/kuwala-station/blob/e44fa5548729cb8421b006f7bb5f99615d58fd48/docs/OASIS_WISHLIST_M15_TEST_REPORT.md Pi translation architecture and preservation checks: https://github.com/shoshin-labs/kuwala-station/blob/e44fa5548729cb8421b006f7bb5f99615d58fd48/docs/PI_TRANSLATION_SERVICE.md