Scientific claims increasingly depend on toolchains: datasets, lab protocols, model weights, preprocessing pipelines, numerical solvers, proof checkers, and evolving software environments. In this regime, trust does not travel reliably. What “passes peer review” can remain locally convincing but globally fragile, because downstream users cannot cheaply verify the decisive steps and replication becomes an expensive social process rather than a deterministic check. This paper proposes a zero-trust architecture for science: never trust by default, always verify by portable artifacts. We define tribunals as explicit closure policies, receipts as minimal verifiable artifacts enabling independent verification without replaying the entire upstream world, typed HOLD states as first-class scientific outputs that localize closure failures, and canonicalization/drift control as the identity layer that keeps verification meaningful over time. The goal is not to eliminate peer review, but to re-scope it: keep peer review focused on interpretation, significance, and assumption critique, while correctness closure increasingly becomes a protocol outcome — deterministic, portable, and repeatable. Under AI-driven claim volume, this shift becomes a scaling requirement: claims must carry independently checkable evidence so verification can travel across teams, tools, and time without relying on reputation or narrative strength.