Don't Merge Carol: Cross-Inhibition and Protected Dissent in Multi-Agent Language Systems

This preprint introduces the Latent Resonance Loop, a multi-agent language system in which agents coordinate through a shared continuous-valued latent buffer rather than text. The project is built in two phases. Phase 1 studies a text-level blackboard system with cross-inhibition and a structurally protected dissenter named Carol. Phase 2 moves the same idea into latent space: agents encode hidden states into a shared buffer, resolve conflicts with a TIES-inspired merge rule, and update the buffer through damped fixed-point iteration. The main technical result is that, after isotropy calibration, the system produces collective latent states that are consistently distinct from those of individual agents. At GPT-2 scale and across three 7B model families, the converged buffer decodes to tokens absent from every individual agent’s top predictions, with measurable divergence from each contributor. The paper also documents an important failure mode: early codec results that looked strong under cosine similarity were largely artifacts of transformer representation anisotropy. The downstream usefulness of the collective state is more limited. Reranking results are mixed overall, though prompts with genuine strategic tension show a clearer advantage for the collective representation. The contribution of the work is therefore not a broad performance claim, but an empirical demonstration that protected dissent and latent-space conflict resolution can produce emergent shared representations in multi-agent language systems. This record contains the preprint manuscript for “Don’t Merge Carol: Cross-Inhibition and Protected Dissent in Multi-Agent Language Systems.”