Context as Architecture

Context as Architecture: Full-Corpus Grounding, Orchestrated Agentic Development, and Compliance-by-Design in a Production AI Assistant Platform

Abstract

Enterprise adoption of large language models is constrained by three problems that are usually treated separately: factual reliability, development throughput, and regulatory conformity. This paper reports a longitudinal design-science case study of a single production artefact — Lumina, a multi-tenant AI assistant platform — across two complete generations built fourteen months apart in capability terms and seven months apart in calendar terms.

Generation 1 (built December 2025, launched 5 January 2026) was produced by one developer, one coding agent, one context window, in a single build cycle of approximately thirty days, and reached production with 274 commits and roughly 100 accounts. Generation 2 (cut over 31 July 2026) was built by the same developer using an orchestrated multi-agent harness, and delivered channel federation across five surfaces, a full teams-and-roles authorisation layer, supervised vendor support access, an analytics subsystem, end-user data-subject-rights tooling, and a complete EU data-residency migration — in a single month, taking the repository from 274 to 867 commits and the automated test suite past 5,700 assertions.

We formulate and evaluate five research questions. RQ1 asks whether a context-engineering discipline developed for small projects survives at platform scale; we find that it does, and that its failure modes migrate from the code into the documentation-to-artefact propagation layer. RQ2 asks what can honestly be claimed about hallucination in a retrieval-free, full-corpus-in-context architecture operated in production for seven months; we argue that the correct claim is structural rather than statistical, we decline to assert a measured zero rate, and we explain at length why the absence of reported grounding failures is a floor and not a rate. We report one operationally significant result: across the period no client reported a fabricated answer, while clients did report insufficient answers, traced in every case to under-provisioned corpora — which is the failure mode a refusal-first grounded architecture predicts. RQ3 asks whether the observed seven-month advance in agentic coding capability was causally necessary for Generation 2; we find strong support for a joint necessity thesis — neither frontier model capability nor context discipline was individually sufficient — and we present a natural experiment in which the release of a frontier model on 24 July 2026 coincides with the eight most architecturally consequential days of the build. RQ4 asks whether an SME can build an AI system that is conformant-by-design with Regulation (EU) 2024/1689; we find that it can, that the artefact met the Article 50(1) transparency duty nine days before its 2 August 2026 application date, and that the residual conformity failures cluster almost entirely in the lag between shipped code and the legal instruments that describe it. RQ5 dissects twenty-seven production capabilities and derives the enterprise requirement each answers.

We abstract eight design principles from the case and state the threats to validity of a single-case, self-reported study with unusual directness. The paper’s principal contribution is not the artefact but the demonstration that documentation-as-executable-context, adversarial cross-family model auditing, and assertion-to-test conversion together constitute a reproducible method for building regulated AI systems at a scale that conventional resourcing would price out of reach for a European SME.

Keywords: retrieval-augmented generation; long-context language models; hallucination; agentic software engineering; context engineering; EU AI Act; Regulation (EU) 2024/1689; GDPR; design science research; multi-tenancy; compliance by design

AI coding agents Lumina RAG