Kuwala / Oasis development checks — 13 October 2026 Oasis is running on Kuwala hardware in the lab. The complete portable station still needs assembly and combined power, charging and Wi-Fi tests. A first local partner trial has yet to be agreed. Current library Oasis Knowledge has 123 published, searchable PDF documents across seven libraries and 39 sections. The private lab catalogue holds 125 records; two encrypted originals remain on hold. The new visitor collection was published on 11 October. Individual PDFs may belong to several libraries. The original files and source details are retained. Questions and follow-ups under sustained load On 13 October, the normal deployed service was exercised for 21 minutes 45 seconds, from the first question to the final completed response. Seven rounds each submitted four initial questions, then four follow-ups using the genuine answers from that round. The questions covered the water cycle, textbook study methods, forage-seed storage and irrigation energy. This was a fixed workload in the lab, not a field trial or all-night test. All 56 requests returned HTTP 200 and an explicit terminal completion. There were 49 local answers and seven explicitly partial answers. The seven initial energy answers were partial after answer validation. No HTTP/stream failures, service restarts or safety stops were observed. This is a functional stability result, not 56 accurate or complete answers. The service admits two active question workflows and four waiting requests. The answering model retains one native generation slot: overlapping workflows do not mean two answers are being generated simultaneously. The middle response time (median) was 61.48 seconds, including waiting. The interpolated 95th percentile was 105.71 seconds. Median queue waiting was 12.17 seconds. These times describe this tested workload, not every question or arbitrary community traffic. What the answer review found All 28 follow-ups retained their own conversation topics. Review still found material limitations: - Farming answers lost species qualifications around forage-seed moisture targets and did not resolve conflicting guidance about drying in sunlight. - Initial energy answers remained partial. Their follow-ups overinterpreted electricity tariffs and conventional-fuel guidance. - Some learning follow-ups attached chapter-summary advice to a passage about SQ3R which did not establish that advice. Source metadata checks passed for 130 evidence records. Of those excerpts, 120 matched current reader text exactly; ten were unverified by the bounded reader check. An unverified match is not proof that the original is absent or different. This HTTP check did not observe actual model source reads; stronger provider-read checks belong to the separate private candidate trial. Correct citation handles and exact excerpts do not prove that every claim has been interpreted correctly or attached to the right source. Access while questions were running Catalogue, keyword search, source passages and original-PDF requests through the Pi visitor listener remained available during the load check. The separate public check verified queue positions, cancelling a waiting question and reader navigation. The three completed public-check responses took 35.17, 46.65 and 62.59 seconds; the fourth was cancelled while waiting and was not a completed answer. Original-PDF transport and file format were verified; visual rendering in the browser's raw PDF panel was not confirmed. A genuine completed water answer was translated into French locally on the Pi in 20.13 seconds. It kept its machine-translation label. Response flags do not establish accurate meaning; earlier translation review found wording errors. Fluent-speaker review remains necessary. Software checks and what comes next The recorded software suite passed: 2,944 tests passed, one skipped and 40 subtests passed. These check software behaviour and contracts, not AI answer accuracy or translation quality. Next checks will cover more source-supported questions, missing information, follow-ups and translation. The complete station needs combined electrical, charging, running-time and Wi-Fi checks. A local trial should compare chat and direct reading of the same documents, then follow up on usefulness and the local team's ability to look after the station. Source and further detail The full report describes the method, measured results, separate candidate and deployed checks, accuracy defects and final restored runtime: https://github.com/shoshin-labs/kuwala-station/blob/b458bad/docs/OASIS_PARALLEL_PROCESSING_TRIAL.md The deployed application release in these checks was e95b473fb98ddfbcc62300e504ebb5b11768bb3b. The current models were Ministral 3 3B for answering on the Jetson, Nomic Embed Text for query embeddings on the Pi, and NLLB-200 distilled 600M for experimental local answer translation on the Pi. Web lookup and chat transcript retention remained off. These public notes contain aggregates and findings, not private conversation archives.