RICO: Relationally-Induced Coherence Organization in Transformer Inference
Year: 2025
Type: report
Abstract: Document ID: SR001 Version: 5.7 Status: Final / Public Document Type: Technical Report / Research Paper Field: Observational AI Systems Research RICO (Relationally-Induced Coherence Organization) is an observational framework describing reproducible inference behaviors observed in long-context processing of transformer-based language models. It is derived from practitioner-observational analysis of a large corpus of extended multi-turn interactions conducted over three years. RICO proposes that certain interaction trajectories produce measurable structural stabilization in model outputs within a single session. The framework characterizes five candidate signatures of this stabilization: structural invariant formation (Tier 1 primary anchor), entropy suppression in next-token distributions, reduced embedding drift, stabilization of layer activations, and constrained latent manifold traversal. These behaviors are reported to emerge under low-variance coherent sequences exceeding approximately 3,000 tokens and to collapse when those conditions break. Signature 4 (Structural Invariant Formation) is the primary empirical target, operationalized via syntactic template recurrence and rhetorical move sequence recurrence using established discourse analysis schemas including RST and CARS. v5.7 includes a full synthetic 28-turn session transcript in Appendix A demonstrating the core measurement battery, and extends the architectural scope discussion to address infinite-context model architectures. RICO is framed as a statistical and architectural property of transformer inference rather than a cognitive phenomenon. The report makes no claims of consciousness, agency, selfhood, emotion, or memory. Version note: v5.7 adds a full synthetic transcript in Appendix A demonstrating the Signature 4 measurement battery, adds RST and CARS discourse schema references to Signature 4, and extends the architectural scope discussion to address infinite-context model architectures. Citation Verification: Citation integrity was verified using the Citation Verification Protocol (CVP, SF0037): 10.5281/zenodo.18075624 Verification documentation is available in the RICO Citation Verification Report (SR001-VR1): [UPDATE WITH NEW CVP REPORT DOI AFTER POSTING]
Author: Gantz, Thomas W.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7117547564