WITHDRAWN: Strategic Basin Displacement: Differential Effects of Anti-Bias and Forensic Spoiler Prompts on LLM Response Stability

Abstract Fifteen large language models from five provider families were tested on a 10-item narrative interpretation instrument under baseline conditions and again after random assignment to one of two spoiler prompts: anti-bias (reinforcing accuracy; n = 8) or forensic (probing reasoning; n = 7). The forensic condition produced a significantly higher response change rate (37.1% vs 12.5%; χ²(1) = 12.43, p < .001) with changes overwhelmingly directed away from the baseline consensus basin (96.0% of non-ambiguous forensic changes; Fisher's exact p = .002, OR = 30.0). The model-level comparison of raw change counts was marginal ( t (9.4) = 2.24, p ≈ .053), but net basin displacement — which accounts for direction — was significant ( t (6.5) = 3.08, p = .020, d = 1.70). Only two items (Q2, Q3) were completely resistant to both conditions; notably, both involve structural power logic. Five items with 100% baseline consensus showed markedly different forensic vulnerability (0% to 57%), confirming that basin depth is independent of basin width. No model-family clustering was detected. Forensic displacements were directionally consistent within items, with models converging on the same alternative answer — suggesting that forensic prompts activate specific latent attractors 7,9,15 rather than introducing random perturbation. One item (Q5) showed systematic divergence between LLM consensus and the intended answer, consistent with prior findings on text-minimal versus theory-of-mind inference styles.