Overtuning by Metaprompt: A Non-Parametric Control Axis for LLM Behavior

Large Language Models (LLMs) are predominantly controlled through parameter-level optimization, with fine-tuning often treated as the primary mechanism for behavioral alignment and specialization. While effective in many contexts, fine-tuning alone operates over globally distributed training data, leaving inference vulnerable to coherence loss when reasoning spans distant or weakly related knowledge regions-an effect frequently associated with hallucination phenomena. This paper argues that fine-tuning should not be regarded as the sole legitimate axis of behavioral control in LLM-based systems. We introduce overtuning, a non-parametric control paradigm that governs model behavior externally through an orchestration layer, without modifying model weights. Central to this approach is the concept of a metaprompt: a software-mediated control artifact responsible for dynamically assembling, constraining, and injecting context-aware prompts into one or multiple LLMs based on persistent memory, pre-registered policies, and execution-time conditions. We present a minimal executable framework in which overtuning reduces effective inferential distance by anchoring generation to dynamically maintained contextual and mnemonic pillars, thereby mitigating hallucination risk through constraint rather than additional learning. Beyond contextual grounding, the framework supports ethical enforcement, role scoping, behavioral routing, and auditability via specialized constraint modules operating at the orchestration level. At the same time, we explicitly address a critical risk of overtuning: excessive internal coherence may lead to epistemic enclosure, where systems become circularly self-consistent and insufficiently exposed to external reference signals-a phenomenon analogous to Plato's Cave. We argue that this risk does not invalidate overtuning, but instead defines concrete governance requirements, including de-circularization mechanisms, external refresh policies, and formal system-level invariants. By formalizing overtuning as a distinct control axis and articulating its architectural benefits and limits, this work contributes a governance-oriented alternative to parameter-centric alignment, with direct implications for explainability, auditability, and the design of controllable AI systems.