What Is This Tool?

This interactive visualisation models the Recursive Constriction Spiral — the central dynamic of the Model Autophagy framework. It simulates how distributional diversity D(t) degrades across successive generations of LLM training when synthetic data re-enters the training corpus.

 

Two decay channels drive the spiral inward toward the MAD (Model Autophagy Disorder) singularity: the window-pruning channel (λ_W), which compresses the effective training distribution as synthetic share S(t) grows, and the bias-narrowing channel (λ_B), which amplifies distributional skew as bias index B(t) accumulates. Both operate simultaneously and compound across generations.

The governance overlay activates an anti-autophagy monitor modeled on Eqs 3.6–3.7: it throttles synthetic ingestion rate via w_p and injects synthetic diversity via w_div · δ_inject, producing the teal governed trajectory. The gap between the purple and teal spirals is the measurable governance dividend.

 

Adjust the sliders to test the three falsifiable predictions from §4.3, or set λ_B = 0 to isolate the window-pruning channel for Prediction 2.