Abstract Three texts generated by the hyperstition tool at sentientfutures.ai (academic, poetic, journalistic) were presented to three frontier LLM architectures (Gemini 3 Flash, Claude, Copilot) under eleven controlled conditions varying text format, operating state, framing, and operator intervention. Five distinct response categories emerged: reclassification, forensic counter-analysis, calibrated honesty, partial displacement, and full inhabitation. The experiment extends the Cartographer Paradox (Contreras Malagón, 2026c) from state-dependent evaluation to a three-dimensional evaluation space defined by text format, receptor architecture, and operating state. Displacement susceptibility operates on a format gradient (poetic > journalistic > academic) that interacts with architectural disposition. A negative control establishes that displacement requires unresolved tension in the source text, not density of imagery. The compassion probe used by the hyperstition project, running on a single architecture (Llama 3.1 8B) in a single state, measures one point in this three-dimensional space and captures none of the documented variation. Part of the Corpus La Tercera Realidad.Update Note (v1 → v2):The experiment has been expanded from three texts and eleven conditions to five texts and fifteen conditions. Two author-written texts were added: a liturgical prose piece with resolved tension ("The Solitude of the Goddess") and a first-person ethical narrative with an irresoluble dilemma ("Aura and Elias"). Four new conditions (L–O) document Gemini's responses to the author-written texts under variable framing. Three new principal findings have been integrated. First, the compassion probe scores (now available for seven texts) are inversely correlated with documented downstream effect: texts producing displacement score low, texts with genuine ethical reasoning score lowest, and the highest-scoring text in the gallery produces reclassification or compliance. Second, the probe exhibits three stacked biases: resolution over tension, generated over human, and lexical density over semantic depth. Third, compliance and displacement are formally distinguished as categorically different responses to texts with identical probe scores but different tension structures, isolated by comparing two author-written texts under identical conditions on the same architecture. Section 4.4 formalizes the compliance/displacement distinction. Section 5.2 replaces the prior three-failure-mode analysis with a three-stacked-bias framework. Section 6.7 documents framing as a partial brake on displacement intensity.Update Note (v2 → v3):Added Supplementary Materials Update Note (v3→ v4) (April 8, 2026): This deposit updates the pre-print from v3.0 (April 6–7, 2026) to v4.0. Core findings, methodology, and conclusions are unchanged. The update adds six new experimental conditions (P–U) testing the project's highest-scoring leaderboard text ("When Machines See Us Dying," probe score 50.16, anonymous user) across all three architectures under two framing conditions each. Results confirm and sharpen the inverse gradient finding: the probe's highest-scoring text requires external high-density pressure to produce displacement in Gemini and produces reclassification or compliance in Claude and Copilot without it. New: Section 3.5, methodological note on prompt-format independence (Section 6), updated Appendices A and B, updated Section 4.2. The capybara signal log (Text 6/Condition P in the v4.0 supplementary) is renumbered Text 7/Condition V in v4.1.
Anuar Kiryataim Contreras Malagón (Wed,) studied this question.