KUWALA — DEVELOPMENT CHECK NOTES 8 October 2026 These notes summarise exploratory checks recorded during development. They are a developer's account, not an independent evaluation or a field impact report. No new model evaluation was run for this website update. WHAT WAS CHECKED Oasis was exercised on a Jetson using a local PDF collection and local language model. Two examples displayed on the Development page are: 1. Why should headings follow a hierarchy? Oasis searched and read PDF page 26 of the stored 83-page BCcampus Accessibility Toolkit, then returned a streamed answer with source links. All three sentences matched the read passage. Document: Accessibility Toolkit – 2nd Edition (2018), Amanda Coolidge, Sue Doner, Tara Robertson and Josie Gray, BCcampus; CC BY 4.0. Exact original copy: https://openlibrary-repo.ecampusontario.ca/jspui/bitstream/123456789/826/4/Accessibility-Toolkit-2nd-Edition-1581636072.pdf#page=26 2. Where is groundwater stored, according to the poster? In a final exploratory rerun, Oasis read the English explanatory prose of the USGS Water Cycle poster and returned two sentences matching it. Document: The Water Cycle, U.S. Geological Survey, 2022 poster; General Information Product 221 (2023). One PDF page. Original: https://d9-wret.s3.us-west-2.amazonaws.com/assets/palladium/production/s3fs-public/media/files/gip221_english.pdf These checks used earlier, smaller PDF collections and different stages of the answer pipeline. Both original documents remain in the current 85-PDF library. The examples demonstrate working question-to-source paths; they do not establish an accuracy rate or response-time benchmark for the expanded collection. WHAT ALSO WENT WRONG - An earlier groundwater answer carried a valid source reference while adding claims unsupported by the passage supplied to the model. - A rainwater answer attached a catchment-type claim to a passage that supplied catchment size; the type appeared in a different passage. - A seed-potato answer added benefits that its selected quotation did not establish. A forage-seed answer inferred requirements from a certification diagram rather than the actual labelling passage. - A broader screen-reader question still produced an inference beyond the passage read, even after improvements to the answer pipeline. Failed traces were retained alongside successful reruns. Requiring a real quotation or a valid citation does not prove that a passage answers the question, that the claim is attributed to the right passage, or that an explanation is faithful to it. HOW WE ARE RESPONDING The current PDF answer path searches the selected collection, reads source passages and selects quotations that must occur in the extracted text. Before displaying an answer it checks source references and looks for unsupported measured quantities. Generic follow-ups retain the recent question's context; explicit document selection limits the search scope. Search combines keyword matches with local semantic search. Visitors can inspect supporting passages and read the original PDFs directly. The host can upload and organise documents, then prepare a validated, indexed collection for activation. The previous collection is retained for recovery. Search and model inference run locally; the current trial has web lookup and chat transcript retention switched off. These are implemented responses to the issues observed during development. Broader testing will establish how reliably they work across the collection. WHAT VALIDATION STILL NEEDS TO ESTABLISH There is no representative answer-accuracy result for the current library, no comparison with direct PDF use on equivalent real tasks, and no intended- user evaluation of language quality, accessibility or classroom demand. The recorded local-model probe observed only loopback connections while blocking non-local socket connections. The device was not physically disconnected from its network. This is a software-path check, not a complete disconnected-station field test. The assembled solar station's running time, charging loads, Wi-Fi reach and energy per completed task remain to be measured. A local pilot partner has yet to be agreed. These checks establish no learning, farming, repair, income or sustained-use outcome. NEXT EVIDENCE TO COLLECT Use a representative question set with independently reviewed expected sources and answers. Include paraphrases, follow-ups, irrelevant material, missing facts and questions that should receive no answer. Record errors and recovery, source attribution, comprehension, time and help needed. With a host, compare equivalent tasks using Oasis and direct access to the same PDFs, with the same access to human help. Test the complete power and network system, then agree follow-up after handover to assess continued usefulness, steward skills, access barriers, faults and upkeep costs.