The Index-Time Context Envelope: A Theoretical Model of Context Propagation in Chunk-Level Retrieval

This paper proposes the Index-Time Context Envelope (ITCE) as a unifying analytical lens for an architectural pattern that is implicit across modern information retrieval techniques. Modern retrieval systems operate increasingly on chunk-level representations of documents, supporting passage retrieval, snippet generation, and synthesis by AI assistants. A widespread assumption follows from the visible behaviour of these systems: that chunks function as independent units of retrieval and selection. This paper challenges that assumption. The paper argues that scalable chunk-level retrieval is expected to rely primarily on precomputed contextual signals attached during indexing, and that the apparent independence of chunks at query time is best understood as conditional on those signals. The Index-Time Context Envelope is the theoretical construct describing the precomputed contextual signals associated with chunks — including signals derived from the chunk's source document and any broader containing structures — read alongside each chunk during query-time scoring. The envelope is a theoretical abstraction over heterogeneous mechanisms (metadata, embeddings, summaries, features), not a discrete implementation artefact. The paper formalises a central thesis — that chunks are evaluated as independent candidates at query time because they are made context-dependent at index time — under a narrowed reading that excludes later reranking, filtering, and synthesis. A constraint-based argument from latency, scale, and chunk-level retrieval is presented to support the necessity of index-time context attachment under those constraints. A predicted failure mode, intra-page intent conflict, in which incompatible signals within a single document are predicted to degrade the retrievability of every chunk that document contains, is set out as a testable prediction with an operational measurement protocol. Implications for content architecture follow, including the prescriptive principle that a page should carry one dominant intent, in the probabilistic sense defined herein, with secondary intents subordinated and structurally tagged. Falsification conditions are stated explicitly. The paper draws on public technical statements by search-engine engineers (Splitt, Illyes, Canel) and Microsoft principal group manager Madhavan (2026), whose public articulation of Microsoft Bing and Copilot retrieval architecture corroborates the index-time annotation account presented in §4. The freshness-failure pattern reported by Madhavan is folded into the paper as parallel observational evidence for the envelope-contamination mechanism predicted by the model. The paper is positioned as a theoretical contribution intended to provide a defensible analytical lens for system designers and content architects working in modern retrieval environments. It is the chunk-level companion to the ten-gate pipeline framework in Barnard (2026, A Ten-Gate Pipeline Model for Entity Visibility Across the Algorithmic Trinity, https://doi.org/10.5281/zenodo.18735074), addressing the index-time annotation phase (G₅) of that pipeline at chunk granularity.