The Floor Beneath the Filter: Provenance Legibility, Pre-Generative Corpora, and the Limits of Model-Collapse Mitigation

Abstract: As synthetic text saturates the web, the ability to read the origin of data is collapsing. The AI industry treats the collapse as an engineering problem to be solved with filters, digital watermarks, and secure tracking. But engineering cannot manufacture provenance; it is a historical state. The ability to verify origin divides the world’s text into two collections separated by a contamination event. The pre-contamination corpus carries its provenance intrinsically, acting as a depleting reserve. The post-contamination corpus requires adversarial engineering to retain origin certainty. All mitigation strategies—watermarking, detection, and strict screening—are paid reconstructions of a property the reserve had for free. Ultimately, the future of artificial intelligence does not depend on scaling synthetic generation, but on anchoring it to the unalterable record of the time before the imitation began. Keywords: Verifiable Origin, Model Collapse, Synthetic Data, AI Infrastructure, Generative Saturation, Data Custody, Contamination Boundary, Depleting Reserve, Low-Background Steel, Artificial Intelligence, Digital Archives, W3C PROV, Data Lineage, Machine Learning, Temporal Boundary