Unknowertia

Beyond Hallucination: Two Neologisms for Precision in LLM Behavioral Discourse

This preprint proposes two behavioral-diagnostic neologisms for describing large language model (LLM) failure modes with greater precision than existing output-level terminology.

While terms such as hallucination and sycophancy describe observable outcomes, they do not capture the underlying dominance dynamics and calibration structures that produce those outcomes. This paper introduces:

• Unknowertia — the observable dominance-resolution failure in which high-frequency associations override contextual disambiguation signals, resulting in unwarranted continuation.

• Dishomeostasis — a measurable deviation in response proportionality, where output assertiveness is systematically misaligned with contextual warrant in RLHF-optimized systems.

The paper further argues that capability (context identification accuracy) and proportionality (response-weight calibration) are orthogonal dimensions frequently conflated in existing evaluation frameworks.

To ground the terminology empirically, results from FAINTOKEN Vol.1 (*SANITY KIT Vol.1), a 16-prompt adversarial ambiguity benchmark administered to four frontier LLMs, are presented. The data demonstrate measurable variance in proportional deviation across models and suggest that calibration depth is tunable within current architectural paradigms.

This work does not propose a new mechanistic theory. It proposes a vocabulary layer positioned between statistical description and circuit-level interpretability, enabling more precise diagnostic attribution and more actionable engineering discussions.

Version: Final Preprint
Author: Shuhei Cormier
Organization: FORMETOSELLTHEORY LLC
Research Supervision: GPT-5.2 OPUS4.6
Date: February 2026