How Evidence Can Support Commercial Payment Decisions

Most companies already hold more operational data than their accounts payable process uses. Telematics, gate logs, RFID reads, TMS records, EDI feeds and document scans all describe the same physical movements that commercial charges refer to. Turning that data into something a payment decision can rest on takes more than access.

Four properties that make data usable as evidence

• Independence — the source is not controlled by the party issuing the charge. A carrier's own status message is informative but not independent; a facility gate record is.
• Provenance — you can say where a timestamp came from, what system produced it, and when it was recorded.
• Precision — the source measures the point the contract actually names, at a resolution finer than the billing increment.
• Reproducibility — the same inputs can be replayed later to produce the same conclusion.

Reconstruction, not averaging

When several sources describe the same event, the temptation is to average them. That is a mistake: an average is not a fact anyone observed, and it quietly hides disagreement. A better approach is to select deterministically — for example, the median observation, with ties broken toward the earliest — and to record the spread separately.

The spread is the interesting part. A two-minute disagreement between a geofence and a gate kiosk is normal instrument variance. A forty-minute disagreement means something is wrong with one of the sources, or with the assumption that they describe the same event. The correct output in that case is not a number; it is a conflict flag and a request for human review.

Separating interpretation from arithmetic

Language models are good at reading documents and poor at being accountable for money. A workable division of labour is strict: models extract and propose structure from contracts and invoices, that structure is validated against a schema before it is stored, and the financial arithmetic runs in ordinary deterministic code with integer cents. The model never decides an amount.

Recording the result

A verification is only useful if it survives the conversation that follows it. That argues for an immutable record containing the transaction, the obligation, the exact contract rule applied, the evidence sources with their provenance, the reconstructed timeline, the calculation steps, the resulting status and a content hash of the whole body. New facts create a new version; they never overwrite the old one.