A General Methodology for AI-Assisted, Evidence-Based and Cryptographically Verifiable Conditional Payment Systems

Preprint (10 pages). A conditional payment releases money only if an agreed event occurred, but today a third party cannot check afterwards why the money moved. This paper presents a methodology that splits such a system into three planes of unequal authority: custody (holds the funds), a deterministic decision function evaluated over signed evidence and signed reception receipts, and an advisory plane (language model and risk scorer) with no write access to the decision. Each outcome is exported as a signed proof package that an independent program replays offline. We define the evidence model, a bounded condition language and the decision function, and prove release, termination and non-interference properties. A reference implementation (AURIX) runs over a payment provider’s authorise/capture/void operations and a BFT ledger. Testing it with random differential testing (20,000 orders, Python vs JavaScript) and exhaustive single-bit corruption of proof packages (192,880 trials) exposed three defects, which we describe and correct. On PaySim, under a time-ordered protocol, a calibrated logistic regression reaches a test PR-AUC of 0.973 (95% CI 0.967–0.978) and gradient boosting is rejected for instability across seeds. The method establishes who attested an event, not that the event happened. Code, experiment scripts and raw results: https://github.com/joeylukoji/Aurix (research/paper).