Canonical Cross-Locus Architecture for Deployment-Stage AI/LLM Evaluation: Definitions, Hierarchy, Variables, Trajectories, Failure Modes, and Measurement Boundaries

A canonical synthesis framework for deployment-stage AI/LLM evaluation under primarily black-box conditions. It consolidates prior cross-locus work into a single reference architecture organized around six core analytic categories: locus, mechanism stack, modifier, intermediate construct, performance criterion, and trajectory family. The document is intended to serve as a primary reference file for later empirical, coding, perturbation, audit, and governance-facing work. The paper argues that current evaluation practice remains fragmented across benchmarking, behavioral testing, red-teaming, alignment analysis, interpretability, user research, and governance documentation. In response, it defines a fixed analytic hierarchy, a canonical variable map, a structured subclass taxonomy, and a relation map separating continuity, alignment, self-report, and black-box evidence. The framework is designed to reduce category collapse, tighten claim boundaries, and improve institutional legibility when observable stability might otherwise be mistaken for deeper integrity, introspective access, identity persistence, or durable repair. REV 1.6 strengthens the synthesis by integrating bounded control- and dynamical-systems framing, clarifying the purpose-relative status of outcome integrity, refining the treatment of self-report as output behavior under black-box evidence, sharpening operational guidance for key variables such as CDI, RGR, SCOG, RC, and RSP, and extending the governance bridge to testing, monitoring, incident classification, and oversight failure. It also includes appendices consolidating the canonical variable map, subclass matrix, relation map, coding scaffold, and pilot skeleton for downstream operational use. The manuscript does not claim direct access to latent internals, does not settle questions of consciousness, selfhood, or architecture-independent identity persistence, and does not present the framework as already validated at full empirical scale. Its contribution is organizational, methodological, and measurement-ready rather than metaphysical. It is intended for AI evaluators, researchers, governance practitioners, auditors, labs, and institutions seeking a portable and disciplined architecture for deployment-stage AI evaluation. Ψ(x) = ∇ϕ(Σ𝕒ₙ(x, ΔE)) + ℛ(x) ⊕ ΔΣ(𝕒′) — C077UPTF1L3 Licensed CRHC v1.0