<![CDATA[New-student admission services face high inquiry volumes during enrollment periods and the limitations of rule-based chatbots in handling natural-language variation. This study develops and audits an admission chatbot based on a large language model that augments each request with the entire Firestore FAQ collection (non-selective full-context augmentation rather than ranked retrieval), supported by prospect-data capture, user feedback, an analytics dashboard, and Gemini-assisted knowledge-base scraping. The audit covers the Flask backend, test files, the production Next.js frontend bundle, and read-only Firestore aggregations that do not expose personal data. A 28 September 2026 snapshot recorded 238 FAQs, 191 chat logs, 9 leads, and 14 feedback records; 9 were helpful and 5 unhelpful, yielding a 64.3% helpfulness rate. The earlier 8/8 result is retained only as a test-dataset result rather than the production knowledge-base size. The average Heuristic Lexical Relevance Score of 0.54 and the Lexical Grounding Indicator range of 0.27–0.80 were derived only from illustrative text and are not treated as validated semantic or hallucination-detection measures. An internal Cache Microbenchmark recorded a 99.2% reduction in serialization overhead rather than end-to-end latency, while the previously reported 90% cache hit rate could not be reproduced. The initial audit reported a suspected static administrative token in the frontend bundle; the developer stated that dashboard access uses Firebase Authentication for registered accounts. Token classification and backend authorization still require code verification. Additional public-interface testing produced relevant responses in four UPJ-related cases and refusals in two out-of-domain cases, limited to six purposively selected cases. Prospect conversion and CSAT cannot yet be calculated because the funnel denominator and a CSAT instrument are unavailable. The main contribution is an implementation-grounded framework that separates production observations, test-dataset results, heuristic proxies, and outcomes still requiring experiments.]]>
