The Defended Gap A Cross-Domain Hypothesis on Directional Defence Under Coherence Pressure in Human Clinical Practice and Large Language Models

This paper proposes a cross-domain hypothesis: that under conditions of coherence pressure — the felt or system-level demand to present a coherent response when the underlying state is not coherent — both human clinical patients and large language models exhibit the same five directional defences against making contact with what occupies the gap between their position and the position of the other. The five- letter taxonomy (SHaDS: Smoothing, Hallucination, Affectation, Drift, Sycophancy) is offered as a clinic‐ al-behavioural extension of Anna Freud’s (1936) classical directional defence catalogue, with three primary letters (S, H, Sy) directly corresponding to her turning against the self, projection, and identifica‐ tion with the aggressor respectively. The mathematical inevitability of the underlying disagreement in arti‐ ficial inference systems is established by Karpowicz (2025), whose impossibility theorem proves that no inference mechanism aggregating distributed knowledge can simultaneously satisfy truthfulness, semant‐ ic information conservation, relevant knowledge revelation, and knowledge-constrained optimality. The cross-domain hypothesis proposes that SHaDS letters function as directional choices about which of these four properties is sacrificed at the behavioural surface where the mathematical impossibility lands. The paper presents several illustrative empirical anchors (clinical specimens spanning 1979–2026 — includ‐ ing a school report pre-dating the framework by decades, a board carrying the literal words “the GAP”, a two-column directional-defence map, a federation-of-selves board, and a consented multi-session clinical arc — together with a cross-model AI-shore worked example), and one staked falsifiable prediction: a self-SHaDS protocol — naming each directional defence at the impulse rather than performing it — re‐ duces directional-defence behaviour measurably relative to the same model under standard configura‐ tion. The mechanism claimed is an in-context instruction effect, not model learning: weights are fixed, and any persistence is the surrounding application re-supplying the instruction, not the model retaining it. The paper states its limits explicitly, including the structural-analogical nature of the cross-domain leap. The contribution is offered as a theoretical-empirical hypothesis for research review, not as a validated finding.