Dongjin Kim

EvidenceDesk is an evidence-first answering tool for Korean small-business owners. Owners paste in their own documents — grant announcements, tax guides, lease summaries, labor rules — and ask questions in Korean. Every answer sentence is quoted verbatim from the source documents with [n] citations, and when the documents do not contain the answer, the tool explicitly declines instead of guessing. Small local models hallucinate when they generate Korean freely, and a wrong tax or funding number can cost a business owner real money. EvidenceDesk therefore removes the model's right to author user-visible sentences: Meta Llama 3.2 1B Instruct only selects which retrieved passages contain the evidence and which sentences to quote, while deterministic code does retrieval, verbatim quoting, citation, and refusal. It runs fully offline on one CPU with 2GB RAM — no API keys, no cards — so business documents never leave the owner's machine. On a frozen set of 24 Korean questions over fictional sample docs (18 answerable, 6 unanswerable), the pipeline cites the correct document on 18/18 and correctly declines all 6 unanswerable questions. A frozen 3-arm ablation (retrieval-only vs Llama passage-selection vs full pipeline) measures what each layer adds: Llama's evidence selection lifts sentence-level evidence accuracy from 0.889 to 0.944, and the validation layer lifts correct refusals from 4/6 to 6/6. An 8-case counterexample set that drove the filter design is kept as a regression suite, not an untouched holdout. These are small synthetic samples — we report them as measurements, not as general performance claims.