Dimensional Gap in Language Model Compression and Reconstruction

This expanded preprint investigates the Dimensional Gap (DG): the hypothesis that higher-order relational content is less recoverable than matched direct relational content after language-model compression and reconstruction. It integrates earlier experiments across several model configurations with two new studies using the interface identified as Gemini Argon. The earlier rotation-controlled DG-v1.1 study comprised 48 compression–reconstruction chains. Direct relations were recovered in 24/144 instances, compared with 8/144 higher-order instances. However, the study did not meet its pre-specified full-confirmation criterion, indicating a conditional, stimulus-dependent pattern rather than a universal effect. The Argon localization study comprised eight full-source controls and 18 compression–reconstruction chains. Analysis identified cases where explicit source information became underdetermined in the compressed text, followed by reconstruction that supplied an unsupported definite claim. A separate six-chain priority pilot showed no observed improvement from prioritizing recognition or document access; four summaries exceeded the requested word limit. The findings demonstrate why agreement between final answers and the original source is insufficient to establish faithful relational preservation. The paper distinguishes source content, compressed text, reconstruction narrative, and final answers, while reporting coding disagreements, protocol amendments, and budget violations. Version 1.2 presents behavioral evidence and methodological limitations. It does not establish a universal DG effect or an internal neural mechanism. Studies using open-weight models and controlled activation interventions are proposed as future work.