Keeping the Model off the Numeric Path: A Voice-First Inventory Ledger for Small Indian Retail in Which Language Models Classify but Never Compute

Natural-language interfaces for accessing organizational data usually place a language model between the user's question and the number they receive. Whether by generating a query or by constructing an answer from retrieved information, both approaches make every displayed number dependent on the behavior of the model, and even with state-of-the-art models, execution accuracy on real-world databases remains far below that of humans. When a wrong number is used without being noticed, as when a store owner orders more stock based on a fake quantity, probabilistic accuracy is not enough to ensure safety. This paper introduces Retail Genie, a voice-first inventory and sales system designed for small Indian retail shops. It is based on a contract that separates the capture of data from its analysis. The model identifies what the user wants to do, maps the user's spoken words for items onto the shop's own closed list of names, and transcribes the numbers the user spoke or that were printed on a bill. At no point does the model have access to stock levels, prices, or revenue figures, nor is it ever asked to perform calculations, aggregations, or rankings. All analytical figures are generated using SQL over an append-only ledger, and any name returned by the model must be verified against existing rows before stock can be moved. On an author-constructed set of 39 utterances, the on-device parser answered 74.4% of them with no model call, all of them correctly, and never placed a declared line on the wrong shelf. On the same set, the deployed model endpoint correctly classified 38 of 39 utterances, resolved all 28 item names and all 15 declared quantities, returned no name outside the shop's list, and had a median response time of 1.5 s. None of seven adversarial inputs produced an analytical figure; however, one of them, an item name that ended the delimiter around the list, caused the model to return a sale quantity that the user had not stated. Removing angle brackets and line breaks from item names did not prevent this; a deterministic check that each quantity returned by the model was actually spoken does. All 490 automated tests pass, including row-level security tests run against PostgreSQL. Evaluation on unrehearsed shopkeeper speech and a comparison with a baseline remain to be completed.