Arkhe e Catedral: Documentação Explicativa em Duas Camadas do Ecossistema de Infraestrutura Verificável para Agentes de IA

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Documentation for autonomous AI agent infrastructure tends to bifurcate into two equallyunhelpful extremes: technical jargon, inaccessible to managers, regulators and citizens, andgeneric popularization, which strips away the precision that would make the materialusable by engineers. This paper presents an application of two-layer explanatorydocumentation — lay and technical — to the Arkhe–Cathedral ecosystem, a verifiableinfrastructure project for AI agents (a containment kernel, decentralized identity,cryptographic provenance and invariant-based governance) recorded in an internal corpusof roughly twenty successive definitions. The method combines the explanation quadrantof the Diátaxis framework, the plain-language principles of ISO 24495-1, and progressivedisclosure, while preserving the project's own epistemic discipline: no claim without a verifiable origin, with an explicit distinction between stipulation and discovery. As a result,the nine auditability contracts (Z1–Z9), the foundational concepts (zero-token architecture,burn-after-use sessions, crypto-erasure) and the structuring metaphor of the Cathedral —the criterion whereby each revision must leave the system more auditable than it found it— become legible to both audiences without cross-layer contradiction. Limitations arediscussed: the corpus's naming fragmentation, the gap between specification andimplementation recorded by the project itself (the form-without-function diagnosis), andthe residual risk that analogies create apparent understanding. The paper concludes that theproject's evidence discipline is itself the most valuable content of the documentation — andthat documenting it in two layers is what allows non-specialists to join in demandingfalsifiers.