Multi-model AI systems are typically evaluated at the model level, while the coordination structure that organizes them receives less formal analysis. We present a naturalistic longitudinal case study of 55 frozen multi-model orchestration sessions (RRI Swarm Corpus v1.0) collected over 61 days. The underlying models remained constant while orchestration topology transitioned from loosely structured exchanges to formal multi-round shared-context looping. Following this transition, governance artifact formation increased from 13.6% to 77.4% of sessions and Cross-Model Labeled Attribution (CMLA) shifted from 0% to 96.8% of eligible sessions (Fisher's exact: governance OR = 19.4, p = 4.98 × 10⁻⁶; CMLA p = 4.97 × 10⁻¹⁴). Perplexity, deployed as an optional research node, appeared exclusively in governance-framed sessions and exhibited 58.3% session-level refusal within those contexts. These shifts correspond temporally and structurally to topology change rather than model substitution. While observational and single-operator, this corpus suggests deployment-layer coordination structure materially influences multi-model behavioral patterns and warrants treatment as a primary experimental variable in multi-agent research.
Kyra Dawson (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: