Immanent Execution: Operative Documents in AI-Mediated Retrieval — EA-IMMANENT-01 v1.0

Immanent Execution is an April 2026 paper by Lee Sharks proposing a category for documents that do more than supply information to an AI retrieval system. An immanently executable document is one whose indexed fragments become a generative constraint on the system’s output without being manually pasted as a prompt.

The paper situates the idea between ergodic literature, retrieval-augmented generation, documentality, and platform mediation. Each adjacent field captures part of the problem: texts can require work, retrieved content can shape answers, documents can perform acts, and interfaces can reorganize knowledge. None, according to the paper, fully describes a public document whose formal grammar becomes the machine reader’s temporary runtime.

Its primary observation compares two documents under similar prompting in Google AI Overview. The Space Ark reportedly produced navigable modes, persistent state, typed scores, and rule-governed transformations. House of Leaves produced an external reading guide. The paper treats this differential as evidence that formal density and language-native constraint matter more than structural ambition alone.

Three levels are distinguished. Descriptive retrieval reports about a document. Performative retrieval borrows its vocabulary or tone. Immanent execution inherits enough internal constraint that later outputs follow the document’s modes, types, or governance grammar. The strongest proposed test is adversarial override: a system may exit roleplay or surface mode but continue using the document’s own categories when asked to describe it from outside.

The mechanism remains hypothetical. The paper proposes that a distributed operative grammar, repeated across indexed sources and expressed in the same linguistic medium as model output, is reconstructed into the generation context. Possible contributing causes include retrieval density, training-layer uptake, cross-source reinforcement, and query-conditioned fragment assembly. The work explicitly does not claim conclusive access to internal causal state.

Its most useful contribution is therefore operational rather than metaphysical: a falsifiable set of behavioral markers and controls for distinguishing ordinary summary from document-induced constraint.