"qsociety team on AMD AI League Hackathon"

A retrieval-augmented question answering container for AMD GPUs that treats every answer as a claim that must be proven from the corpus before it is written. A resident Qwen3-VL-8B-Instruct (bf16, greedy, transformers on the mandated ROCm base) serves a unix socket; each query is a thin stdlib client that always writes a valid JSON file within budget. Indexing walks the corpus defensively (unreadable, encrypted, binary and empty inputs never stop it) and turns PDFs, DOCX tables, multi-sheet XLSX, CSV, logs, Python literals and images into atomic records, with RapidOCR plus VLM transcription for images. Retrieval fuses BM25 with an exact rare-ID index and follows ID chains across files. Withdrawn and superseded revisions are resolved before the model reads anything. Cross-sheet joins link a part's lead time or cost to the product that uses it, and PDF family tables become one record per product. A deterministic verifier checks the value against a token-boundary span in an eligible record, the question's product, quantity and unit, code literals over stale comments, and false premises; a chain guard re-reads log-to-ticket chains when a topical decoy wins. Finally the model answers a YES/NO warrant question over the evidence, which refuses unanswerable questions instead of guessing. Citations come from record identity and chain necessity. Validated on the real model against the sample (200/200, A100 and Radeon Pro W7900D under ROCm), 13 generated adversarial suites with held-out seeds, and a 300-file corpus with 61 images: p95 5.4 s per question, 22.5 GiB peak VRAM.