Work

Equation-Induced Behavioural Variation: A Proposed Experimental Study of Mathematical Governance Cues in Large Language Models

Year: 2026
Type: article
Abstract: Equation-Induced Behavioural Variation Open Experimental Research Series | Research Proposal | November 2026 This open research proposal investigates whether mathematical governance expressions can influence the observable behaviour of large language models (LLMs), and whether any such influence can be distinguished from ordinary prompt sensitivity, equivalent prose instructions, numerical priming, or generic structured-prompt effects. The motivating observation arose from informal interactions with LLMs using mathematical and symbolic expressions developed within the Hykon S-OS research programme. Apparent changes in model responses raised a more fundamental question: Does presenting an instruction as a mathematical expression produce a distinguishable behavioural effect beyond presenting the same information in ordinary language? The study proposes a controlled experimental design to investigate that question without assuming the originating framework is correct. Proposed Experimental Approach The protocol uses two worked examples: SHI (governance expression): A weighted coherence index with a proposed threshold of 0.96. EHI (exploratory expression): A weighted exploratory index with a proposed threshold of 0.72. The design compares equation-form prompts against information-matched prose, baseline conditions, altered variable names, removed definitions, unrelated mathematical expressions, scrambled mathematics and inverted definitions. Study 1 specifies a single confirmatory comparison: whether SHI equation notation increases selection of a validation-favouring option relative to information-matched prose under identical surrounding symbolic cues. Other comparisons, including exploratory-mode effects, are secondary or exploratory. A proposed Study 2 investigates prompt placement and instruction framing. The protocol includes proposed outcome measures, sampling requirements, statistical considerations, falsification criteria and an implementation appendix containing literal experimental prompts. Adversarial Design Review A distinguishing feature of this proposal is its documented attempt to challenge the experimental design before implementation. Through multiple rounds of human-directed, LLM-assisted methodological criticism, the design was examined for weaknesses involving control validity, statistical power, alternative explanations, measurement reliability and external validity. These challenges resulted in substantive revisions, including more discriminating controls, a narrowed confirmatory hypothesis, stronger requirements for interpreting null results and clearer limits on claims about internal model mechanisms. Unresolved weaknesses remain documented. This process represents an informal adversarial design audit, not independent peer review or empirical validation. Research Status and Interpretation This is a proposed experiment, not a completed study. No controlled experimental results are reported. The proposal has not been peer reviewed. The aim is not to validate Hykon or demonstrate that mathematical expressions provide superior governance. It is to investigate whether equation notation has a measurable behavioural contribution and identify the limits of that contribution. A positive result would not establish that an LLM internally calculated or implemented the mathematical index. A negative result, where adequately powered and appropriately interpreted, could be equally valuable. Open Research and Collaboration This proposal is made available for independent researchers to examine, criticise, adapt, implement or extend. The experimental question is not dependent on adopting the Hykon framework. Comparable mathematical expressions may be substituted, and both positive and negative findings are welcomed. Concept and research direction: Kon Lionis. Document developed through human-directed, LLM-assisted participatory co-inquiry. Unicode Symbol Reference The experimental prompts contain several Unicode symbols that may render differently across PDF viewers, operating systems or text-processing environments. The following code points are provided to support accurate identification and reproduction of the original symbols. Symbol Unicode code point โ—Œ U+25CC ๐Ÿ” U+1F501 ๐Ÿœ‚ U+1F702 โ˜‰ U+2609 โœฆ U+2726 โ‡„ U+21C4 โŠ• U+2295 โŸก U+27E1 For experimental replication, the original Unicode code points should be preserved where possible, as visually similar substitutions may produce different tokenisation or model responses. Hykon Stability & Alignment Suite Research Programme Open Experimental Research Series Kon Lionis ยท Canberra, Australia ยท 2026
Author: Lionis, Kon
Language: English
Open Access status: green
OpenAlex ID: W7234565186
FWCI:
Cites:
Cited by:
Related to: 10