The Constitutional Observability Platform

Admissibility engineering and agent-facing governance require more than rules: they require infrastructure that can mediate consequence-bearing operations, generate bounded evidence, evaluate evidence health, expose uncertainty, and preserve a reviewable relation between specification, configuration, realization, effect, and revision. This paper proposes the Constitutional Observability Platform (COP) as that infrastructure pattern. COP consists of eight cooperating roles: a governance substrate (UGK), evidence health monitor (EHM), governed scenario runner (HeadlessRunner), constitutional test runner (CTR), codex-to-implementation projection layer (CRP), agentic interface surface (AIS), evidence verification surface (EVS), and situated awareness surface (SAS). The roles are architectural contracts, not claims that a named implementation conforms merely by existing. The central unit of observability is a scoped claim: subject, proposition, identity, valid time, transaction time, observer, method, evidence, confidence, and known remainder. Receipts may bind such claims and make later recomputation possible, but hash-chain integrity does not prove that an effect occurred, that all relevant pathways were mediated, or that the observing mechanism was correct. COP therefore separates evidence production, effect observation, verification, health assessment, and epistemic projection. The paper also generalizes the display-fidelity problem. A domain model, widget tree, accessibility tree, rendered pixels, user interaction path, and human interpretation are different surfaces. A self-issued fidelity receipt comparing two correlated internal representations cannot prove all of them. Display assurance requires declared correspondence contracts, independent observers appropriate to each layer, negative controls, and an explicit unobserved remainder. The source reports prior clean-room builds and application metrics, but supplies no exact artifact-and-evidence package in this version-family head. Those values are therefore retained as verification targets rather than promoted as empirical results. This candidate specifies the evidence required for self-application, coverage, convergence, display fidelity, and cost claims while preserving the architecture as a reversible author-review proposal.