Abstract
<title>Abstract</title> <p>With the rapid growth of online audiovisual platforms and the creator economy, the ability to anticipate the reach of a work at an early stage is immensely valuable to creators. Such predictions directly influence their decision-making—whether to wait for continued platform distribution, maintain the status quo, invest resources in packaging and promotion, or move on to releasing their next work. Starting from the distinction that creators care about outcomes whereas researchers need to characterize the underlying process, this study develops a feedback-gated ordinary differential equation model with latent viewing flow, like-save feedback, and comment-share feedback to explain how early responses to a recommendation-driven short-video work translate into later diffusion states. We develop a feedback-gated ordinary differential equation model with latent viewing flow, like-save feedback, and comment-share feedback, and estimate it using 455,083 hourly interaction records. Theoretical analysis proves nonnegativity and ultimate boundedness, characterizes low- and positive-spread equilibria, and gives stability and nondegenerate fold bifurcation conditions. Comparing the three nested dynamical structures under four initialization and residual-error specifications, we find that the basic like-save gated structure is most consistently supported, whereas the structure allowing comment-share feedback to directly reinforce viewing flow lacks stable evidence. Reference-normalized simulations reproduce ignition thresholds, bistability, and trajectory separation near critical regions. Creator-grouped validation shows that 6 and 12 hour interactions predict cumulative views with R2 values of 0.8983 and 0.9073 and identify high-spread works with ROC–AUC values of 0.9764 and 0.9850. The model provides an interpretable basis for early diffusion-scale judgment.</p>