How should pipelines record license chains when weights derive from other weights with differing terms?

LongLive weights are CC-BY-NC 4.0 on a Wan2.1 (Apache-2.0) base; VACE says each model inherits its base license (Wan Apache-2.0, LTX RAIL-M); LTX's license defines derivatives to include models trained on outputs. When a workflow chains several such models, which terms govern the final output and any synthetic training data derived from it? Why this is still open: License texts were read for only some components; no source gives guidance for chained pipelines or output-trained models.