Echo Collapse: The Empirical Structure of Conversational Shape

Conversational failures are failures of shape. The original Echo Collapse paper proposed that many assistant failures live in the relational shape of an interaction rather than in any single answer, and named three families of breakdown: over-coupling, task drift, and failed transfer.

This note asks a different question. If conversational shape is measured across many interactions, does the resulting space contain real structure? And if it does, is that structure a property of the conversations, or of the instrument used to measure them?

Several findings emerged. The measured space contains stable, low-dimensional structure. Its dominant dimension — conversational pressure against conversational independence — survives reconstruction through an independent embedding model. The structure does not decompose into the architecture's failure taxonomy. And at least one intervention family moves conversations through the space in a consistent direction, not merely in their wording.

The taxonomy routes action. The structure is something else.

These are exploratory, null-controlled results from a modest corpus, and the sample sizes are stated rather than smoothed over. They do not establish that conversational shape has dynamics, attractors, or stable trajectories. That question remains open. They suggest something narrower and more durable: that the architecture is measuring something more fundamental than the categories it uses to organize corrective action.

Like the architecture paper, this note states results and omits machinery. Signal definitions, thresholds, prototype contents, and calibration details are not disclosed.

Companion to Echo Collapse: A Shape-Based Approach to Conversational Failure in LLM Assistants (Rowe, 2026, https://doi.org/10.5281/zenodo.20278151).