OmNi-Ledger: Deterministic Supply-Chain Control

Pharmaceutical distributors can receive invoices with an incorrect SKU, batch, expiry date, or quantity. Expired goods must be blocked before entering available stock, and a recall must identify remaining inventory and affected recorded dispatches. OmNi-Ledger separates extraction from authority. Gemma reads invoice text through AMD Token Factory and returns structured fields. The app treats model output as untrusted, validates it against a strict schema and approved catalogue, then sends it to deterministic warehouse rules. The model never directly changes inventory balances. Controls include expiry hard blocks at receipt, QC-gated posting, integer quantities, FEFO allocation, scan-verified picking, separation of duties for recalls, warehouse-scoped quarantine, and guards against picking recalled stock. Batch tracing connects recorded supplier receipt to recorded customer dispatch. Inventory movements and audit events use SHA-256 hash chains; a state hash and Merkle-root seal can be verified. These provide tamper evidence, not blockchain or regulatory certification. In a recorded live CLI test on AMD Token Factory, a valid invoice was ACCEPTED and an expired invoice BLOCKED with EXPIRED_GOODS; the gateway identified self-dploy/Gemma-4-31B-IT on AMD GPUs. The project includes a one-click evidence pack with invoice checks, recall evidence, audit data, and exports. The packaged release passed 181 automated tests and has no npm package dependencies. Browser scenarios use fictional local data, not live inference. Built independently in Egypt using Termux and a phone-first workflow, OmNi-Ledger shows how AI can extract information while deterministic rules govern inventory decisions. Traceability is batch-level through recorded dispatch, not patient-level or individual-unit tracking.