The

FourthMatch is a payment-control agent for procurement teams. It stops companies paying for goods that never arrived. Today, suppliers get paid on a “three-way match”: if the purchase order, the goods receipt (GRN) and the invoice agree, the money goes out. But all three are paperwork. If a truck brings 12 boxes and the storekeeper signs for 14, all three documents still say 14, and the company pays for two boxes it never received. Short shipments, damaged units invoiced at full price and rubber-stamped receipts leak money every day, and nobody checks the dock camera footage that already exists. FourthMatch makes that camera the fourth match. When a delivery is unloaded: Meta’s Muse Spark watches the footage and decides what to look for (“cardboard box”). SAM 3.1 segments and tracks every unit, frame by frame. Python merges the tracks and does the counting. The model never does arithmetic. The system then reconciles PO, GRN, invoice and video, and a decision agent applies the company’s own receiving policy. Each payment is approved, part-paid, held, or sent to human review. Every decision explains itself in plain English and quotes the exact policy clause it rests on. For example, §8.4: “when the video and the GRN disagree, the video prevails”. Each decision also comes with: the evidence frames and SAM track IDs, a full trace of the agent’s tool calls, a drafted debit note, an evidence pack (PDF and JSON) sealed with a SHA-256 hash that anyone can re-verify. The same engine also: catches paper-only problems such as price above PO, duplicate invoices and over-billing; checks a storekeeper’s voice remark against the video; lets finance edit a policy clause and replay every past decision against the draft before approving it. Weak evidence is never approved; it goes to a person. On our 12 synthetic cases, verdicts match ground truth 91.7% of the time and every video count is correct. All data in the project is synthetic.