HARMONIC CIRCUIT / MAIN FIGURE EXPLORER

How Information Flows
Through Music and Model Layers

Trace the prediction path to the final tonic chord across real examples, using the original scores, Emergence Orders, and principal-edge weights.

INTERACTIVE FIGURE

Auto-Emergence Principal Information Flow

—

Key
Target
Model
ChatMusician
Edge contribution — —
Emergence Order 1 · earlier later · —

Layered principal information flow from Auto-Emergence The horizontal axis shows music tokens, and the vertical axis shows Transformer layers. Edge color encodes the source token's Emergence Order, while line width and opacity continuously encode edge contribution.

MUSICAL EVIDENCE

Score Mapping for the Same Excerpt

Labels beside notes show each note event's Emergence Order. Labels above chords show the minimum order among the tokens forming each chord event. The gold region marks the target tonic chord that has not yet been provided to the model.

INSPECT

Hover over an edge, token, or score note to inspect exact values.

Auto-Emergence

Starting from the opening-quote anchor of the final tonic chord, recursively selects the highest-contributing cross-token edge from the current exposed pool.