Definition
The fraction of each training or fine-tuning mix that descends from prior model outputs: synthetic generations, model-filtered logs, and model-labeled examples, weighted by generation depth.
The Measurement Gap
There is no metric that tells you the fraction of training data descended from prior model outputs, which means distribution narrowing compounds invisibly across generations.
Formula
TDRF = sum(w_e * depth_factor_e) / total_examples, where depth_factor grows with generations of model descent
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Whether the training loop is narrowing toward its own outputs, and what recursion ceiling each model class tolerates.