Claim-Scoped Relational Evidence Gates for Experimental Computing: A Bounded Reproducibility Study

Experimental computing often combines functional tests, stored-record integrity audits, electrical timing reports, power-cycle observations, and external assessments. These objects support different claims. A record that is internally complete, however, need not establish that the underlying experiment occurred or that its evidence belongs to the claimed device and session. We specify and implement a deliberately bounded claim-scoped relational metadata gate. Five evidence classes require kind-specific fields; witnesses are additionally bound to a claim kind, subject and session, with limited ordering checks for persistence and a declared role-separation check for review. The output is READY_FOR_INSPECTION, INCOMPLETE_METADATA or INCONSISTENT_METADATA, never a physical pass, certification or authenticated proof. We reproduce an exhaustive 15-bit presence study yielding 163,840 claim assessments and 2,160 deterministic structured-record tests covering missing fields, mismatches, empty values and unrelated evidence. An earlier presence-only gate accepts 630 intentionally inconsistent cases. The relational gate rejects all 630 and matches the fixture specification in all 2,160 cases. Additional ordering and role-separation fixtures test narrower conditions. These are self-authored synthetic experiments, not independent corroboration or estimates of real-world error rates. The contribution is an auditable demonstration of conservative metadata-readiness checks and their limits, rather than a hardware qualification or external certification system.