Field-Conditioned Invariance: A Proof-Carrying Framework for Scientific Claims Across Changing Environments

Scientific conclusions are often transported across changes in environment, representation, scale, measurement, instrumentation, population, and computational architecture. A persistent value, similar prediction, or locally valid transformation, however, does not necessarily preserve either the scientific property of interest or the evidentiary warrant supporting a claim about that property. This paper develops Field-Conditioned Invariance (FCI), a candidate methodological integration of Invariant Science and Field Science. It represents the investigation environment as a typed relational description and binds each scientific claim to an admissible transformation, a property-preservation obligation, an evidence-preservation obligation, a reconstruction method, and an explicit failure scope. The central Dual Invariance Gate separates the preservation of the target property from the continued adequacy of the evidence authorizing its assertion. A compositional extension examines whether locally qualified claims remain qualified through successive or reordered transformations, including cases of information loss and observational equivalence. The framework additionally distinguishes initial claim admission, continued survival of evidentiary warrant, and subsequent requalification when new evidence becomes available. A claim-indexed first-warrant-break diagnostic records the earliest documented point at which a previously accepted claim loses adequate support, which need not coincide with an operational failure. Five scientific case studies—NASA-style aerodynamic similarity, universality in critical phenomena, genotype–environment response, transportability of clinical trial effects, and migration of AI agent systems—illustrate different preservation and obstruction structures. A deterministic decimal-rounding demonstration and narrowly constructed trajectory examples illustrate the limitations of end-point-only comparisons; neither constitutes independent experimental validation of FCI. A preliminary documentary inspection of one public AI-agent trajectory is reported separately from controlled evaluation and yields no demonstrated improvement over the original failure-localization annotation. We specify an independently adjudicated benchmark against causal, statistical, provenance, assurance-case, and formal-contract baselines with information-equivalent access. The framework claims neither a universal field ontology nor a new causal identification theorem nor established superiority over conventional methods. Its contribution is a falsifiable, cross-domain proposal for controlling when a transformed scientific claim is entitled to retain its meaning and authority.