Building a deeply AI over a substrate you cannot retrain: the Clara Nexus reference implementation of layered intelligence theory

Layered intelligence theory (LIT) specifies eight design principles for what its author terms a Deeply AI—a cognitive architecture whose parameters and surrounding structure jointly instantiate the preconditions for reorganisation, reflection, and falsifiable learning. The specification is silent on how a builder without write access to those parameters might proceed. This paper reports a working reference implementation, the Clara platform, instantiating those principles over hosted frontier large language models. The reference implementation is the platform as a whole rather than any single surface: four surfaces across three substrate vendors—the conversational surfaces (Clara Nexus chat and Clara Companion) on OpenAI substrate models, the reading and instrumentation layer on Anthropic substrate models, a data-sensing path on Perplexity, and an independent read-only Research seat on Claude Sonnet 5—disclosed in full in §2.4, of which Clara Nexus is the primary surface for the constitutional runtime. This is a reference-implementation and specification-conformance paper rather than a benchmark-driven empirical evaluation, and it makes no task-performance claim. Specification conformance, as reported here, means a traceable mapping from a published specification clause to a named runtime carrier—a typed identifier, a function, or a database table—inspectable independently of the author’s description. The eight principles are sorted, as this paper’s analytic contribution, into those describing what the model’s parameters must do, which are structurally unreachable from a hosted substrate, and those describing what the surrounding architecture must do, which are implementable at the scaffold layer. For each principle, there is a record, a verdict of built at scaffold, partial, or substrate-unreachable, alongside the named runtime carrier grounding it or the gap preventing one, and the surface evidencing it. The core engineering contribution is a constitutional runtime: seven fail-closed checks enforced at prompt-composition time, before any call reaches the substrate, encoded as a single assert Constitutional invariant. Six are typed assertions over a structured candidate move, and one is a pattern-matched textual check whose error characteristics are stated rather than assumed. Perceptual control theory primitives are reported as prompt-level Companion voice discipline rather than as a runtime-enforced mechanism. The empirical evaluation of the deployed system belongs to a companion paper whose eight falsifiable predictions are enumerated in §4.3.