Designing the Real-Data-Anchor Filter Against Model Collapse

Your synthetic-data engine has been running for a year across many slices. A retrospective analysis finds that a specific slice's synthetic pool is now composed as follows: cycle 1 was generated from real images (100% real-anchored); each subsequent quarterly cycle generated new synthetic examples using, in part, the previous cycle's synthetic output as reference/style examples to speed up generation, because doing so produced batches that passed the validation gate faster. By cycle 5, roughly 60% of the slice's active training pool traces back through at least one prior synthetic generation rather than directly to a real image.

1. Explain why "passed the validation gate faster" is itself a warning sign here, not evidence the pipeline is healthy.
2. Describe concretely how you would detect that this slice is at risk of model collapse, using signals available in this pipeline.
3. Propose a specific, mechanical policy change (not just "be more careful") that would have prevented this drift from cycle 1 to cycle 5, and explain the trade-off it introduces.