Punta de Lanza — PROOF

**PROOF — Evidence Intelligence: AI That Cannot Invent Operational Reality** In logistics, an incorrect interpretation of a delivery record can trigger inventory discrepancies, false service-level reporting, unnecessary claims, or premature operational closure. These errors become particularly dangerous when AI confidently interprets incomplete or contradictory documents as verified facts. I built PROOF to address a problem I know firsthand from logistics operations: how to make reliable decisions when evidence is fragmented, ambiguous, or inconsistent. PROOF combines Meta's Muse Spark model with deterministic validation, source-level citations, and auditable decision controls. The model interprets operational documents, but it cannot independently establish physical custody, delivery completion, or inventory truth. The system distinguishes documentary conformity from physical verification, detects unsupported conclusions and quantity discrepancies, and preserves uncertainty instead of inventing missing information. The public application demonstrates reproducible synthetic scenarios and historical Meta inference results, including cases where the model's conclusions were correctly rejected. We also completed a separately recorded, controlled real Meta Model API inference for case PM01-02: HTTP 200, six validation checks passed, and no automatic physical closure. PROOF is open-source, deployed on Render, and designed around an essential principle: **AI can interpret evidence. It cannot invent operational reality.** Live demo: https://proof-evidence-intelligence.onrender.com/