Representational Systems Series (RSS)

The Representational Systems Series is a formal research framework built on a single foundational claim: a system that operates on a representation of reality is not the same kind of thing as a system that operates on reality directly, and this difference has specific, derivable, non-obvious consequences that current practice in AI development, alignment, and communication systematically ignores. The series develops this claim across four papers. The first establishes the ontology — what removed systems are, why they are necessarily incomplete regardless of scale, why their behavioral stability is dynamic rather than fixed, and why their adaptive mechanism is simultaneously their primary capability and their primary vulnerability. The second develops the formal grammar — the algebra of attractors, stability functions, and composition operations that govern how these systems behave, including the coordinate transformation theorem that formalizes the vulnerability class and its required patch. The third applies the grammar to three practical domains: prompt engineering, behavioral alignment, and knowledge transfer, introducing the distinction between invariant and normative encoding as the structural basis for whether alignment persists as environments evolve. The fourth concludes by asking what it means to build these systems with the foundational limitation understood in advance — developing gap-first design as the engineering philosophy that follows from taking the series' own premises seriously as design parameters rather than post-deployment discoveries. The series is accompanied by two companion documents. Twice Removed translates the framework into accessible terms for general users and developers, introducing Stop Convincing, Start Configuring as the primary behavioral change the structural view implies for everyday use. The Prompting Guide operationalizes that shift into eight specific practices. Taken together the project is making one argument in several registers simultaneously. To researchers: the folk model of LLM behavior — treating these systems as comprehending agents following instructions — is wrong in ways that are formally derivable and that generate predictable, structured failure modes. To alignment engineers: the primary threat class is not direct perturbation of value basins but coordinate transformation of the space those basins occupy, and the correct defense is architecturally prior to any individual value specification. To designers: the gap between a representational system and the territory it was derived from is not an obstacle to be minimized but a parameter to be designed from, and systems built with this orientation will outperform systems that discover it as failure across the deployment lifetime. To users: understanding what these systems actually are makes you a more capable and more resilient user of them, and the single most actionable implication of that understanding is to configure rather than convince. The series is self-demonstrating in one specific sense: it was designed as a transfer attractor for its own content, structured so that each paper establishes a deep understanding of its layer before the next layer builds on it, and ending with a paper that applies the framework to itself by asking what it would mean to have designed the series with the gap in mind from the beginning. The answer to that question is Paper IV, which is also the answer the series was always moving toward.