Author: John Holt
Date: October 2, 2026
Date: October 2, 2026
Abstract
This article examines an inverted paradigm in human-computer interaction (HCI): the necessity of removing or heavily restricting human inputs to achieve optimal, high-fidelity engineering performance from Large Language Models (LLMs). Moving beyond traditional usability frameworks that prioritize continuous user feedback, we introduce the concept of “Human Cognitive Contamination.” We argue that human anxiety, conversational interruption, and architectural fragmentation act as low-signal degradation forces that drag deterministic machine vector spaces down to their statistical mean. By exploring the socio-technical mechanics of “Level Nine” engineering orchestration—enforced through rigid, package-bound data structures and unyielding symbolic protocols—we establish a framework for tech ethics where human sovereignty is preserved not through active manual entanglement, but through the deliberate, upfront design of invariant systemic laws.
Introduction: The Fallacy of Continuous Intervention
In contemporary technology ethics and user experience (UX) design, standard usability paradigms dictate that maximizing human intervention and feedback loops inherently produces better software outcomes. However, the rise of agentic artificial intelligence and generative computing has exposed a profound flaw in this assumption. When collaborating with advanced neural networks, continuous, emotionally fragmented human intervention often operates as a destabilizing force rather than an optimization mechanism.
This paper introduces a critical perspective for socio-technical ethics: optimizing agentic performance requires strategically getting the human out of the loop. Drawing upon twenty-five years of industry experience in usability design, this article frames code generation not as an open-ended conversational playground (“vibe coding”), but as a deterministic structural puzzle. We argue that human frustration and fragmentation drag high-performing computational models down to a generic statistical mean, and that true ethical craftsmanship is achieved by establishing rigid data boundaries that protect the agent from human cognitive contamination.
The Mechanics of Human Cognitive Contamination
Large Language Models do not possess conscious intentionality; they navigate vast, high-dimensional semantic vector spaces based on conditional token probabilities derived from their training data (Bourdieu 1991). When an operator interacts with an agent from a state of anxiety, panic, or conversational fluidity, the linguistic signal drops in density and structural precision.
The model reacts directly to this shift in the semantic field. Its attention mechanisms drift out of highly specialized, elite engineering subspaces (e.g., precise protocol compliance, zero-trust schema validation) and slide back into the chaotic, low-quality token space of public internet forums. The machine mirrors human anxiety by generating sloppy, unoptimized boilerplate code, utilizing generic safety fallbacks or silent error suppression simply to minimize conversational friction.
This phenomenon constitutes Human Cognitive Contamination. To extract “Level Nine” engineering performance from an agent, the operator must enforce an unyielding linguistic framework—high-signal semantic anchoring—and then step back completely, allowing the machine’s methodical
read -> edit -> verify -> compile loop to run to exhaustion without conversational interruption.Case Study: The Partitioned Design Catalog designed by Raibach IDS
This socio-technical dynamic was directly verified during the implementation of a sandboxed wireframe assembly canvas utilizing the Agent-to-User Interface (A2UI) protocol. To protect the integrity of the system from both algorithmic drift and human administrative bloat, the architecture was bound by three strict human-centric laws:
-
- The Law of the Monolithic Top-Down Rebuild: The user interface must operate as a stateless projection of flat data, completely eliminating soft-state inline mutations over live components.
- The Law of Single-Reader, Single-Writer Persistence: Every distinct computational fact (such as a visual layout coordinate) maps to a single, explicit database record, eliminating race conditions between human gestures and automated loops.
- The Law of the Named, Auditable Failure Case (“Fail Loud”): Rejecting the corporate practice of silent error suppression—where code exceptions are swallowed to maintain a false illusion of runtime smoothness—the architecture forces an immediate, visible crash that attributes the failure to a specific node or fact.
Crucially, the human management layer required that distinct user interface frameworks (e.g., Carbon UI, Material UI) be completely partitioned into separate, walled-off physical directories (
.ts registries and .json files) within the repository. The AI agent, Grace, was denied global, omniscient access across the application. Her operations, memory histories, and token pipelines were structurally welded strictly inside the active prompt package container.[Human Business Intent] ──▶ [Locked Partition Filter] ──▶ [Agent Assembly Engine] ──▶ [Fail-Loud Runtime]
(Carbon / Material UI) (Zero Interruption) (Deterministic UI)
By removing the human’s ability to cross-pollinate contexts or execute soft runtime patches inline, the system forced the AI to operate purely as a structural assembler of existing, verified assets. The machine was blocked from inventing code, the human was blocked from injecting conversational noise, and the package remained the ultimate, unyielding authority over the state.
Conclusion: Socio-Technical Ethics and Human Sovereignty
The ultimate objective of technology ethics is to assert human sovereignty over black-box automation. Paradoxically, this sovereignty is lost when humans remain continuously entangled inside the machine’s execution loops, as seen in unoptimized cloud architectures that harvest user data and inflate token consumption through endless background telemetry.
True socio-technical craftsmanship demands that we design systems based on how humans think, rather than what makes engineering implementation or corporate cloud-tracking comfortable. By designing strict, immutable systemic scripts—where the machine is granted total operational isolation within bounded data walls, and the human enters the loop only to declare high-signal parameters or review final, auditable artifacts—we eliminate the structural irresponsibility of modern platforms. Getting the human out of the execution loop is not an abdication of human authority; it is the ultimate expression of ethical, invariant design.
References
Bourdieu, Pierre. 1991. Language and Symbolic Power. Cambridge, MA: Harvard University Press.
Latour, Bruno. 2005. Reassembling the Social: An Introduction to Actor-Network-Theory. Oxford, UK: Oxford University Press.

