Evidence-Carrying Operational Claims in Open Systems: Physical Ledgers, Typed Interfaces, and One-Sided Deployment Guarantees

This preprint develops a contract-based framework for evaluating operational claims in open, partially observable, and potentially adaptive systems. Rather than treating safety, service delivery, resilience, or recovery as intrinsic attributes of a system, it represents them as typed, evidence-carrying propositions relative to a declared physical and institutional boundary, environment mechanism, observation history, intervention regime, policy class, shared resources, and finite physical horizon. The framework integrates hybrid path-space models generated by a common modular mechanism; exact physical ledgers that distinguish atomic events from non-atomic finite-variation flows; calibrated observation models and measurement uncertainty; causal identification and transportability; scenario-fixed experiment interfaces; and policy-uniform correspondences between evidence models, computable concrete models, and abstractions. Its principal formal result is a finite-horizon, one-sided deployment certificate that transfers an abstract lower safety value to deployment under partial observation. Statistical coverage over learning datasets, deployment-path probabilities, reconciliation discrepancies, and implementation or abstraction radii are kept as distinct quantities rather than combined into a single confidence score. Claim-sufficient scopes are not assumed to be unique. They are evaluated through a Pareto profile covering completion nonemptiness, query diameter, decision stability, action support, latent sensitivity, and query type. Explicit verdict semantics distinguish accepted claims, contradictions, unsupported refusals, unresolved decision margins, incomparable claims, and invalid records. A machine-readable implementation based on JSON Schema Draft 2020-12 and exact decimal arithmetic checks finite types, relation coverage, physical balance, provenance exclusivity, artifact containment, hashes, and recomputation of certificate quantities. Synthetic examples involving a distributed AI service and human–AI emergency logistics, together with finite counterexamples and reproducible stochastic fixtures, illustrate the framework. The validator does not establish the truth of external evidence, causal assumptions, statistical models, or real-world safety. The work does not propose a universal performance scale, a new causal calculus, or a replacement for formal assurance cases or runtime monitoring.