The Fabrication Ledger: Why the Cost of an AI Hallucination Is Decided Beneath the Model

Abstract Generative models fabricate. The rate can be reduced. It, however, cannot be brought to zero, and the proof is now in the literature. The industry nonetheless continues to treat fabrication as a model defect to be repaired at the model, through better training, retrieval grounding, and human review. The ledger of documented incidents from 2023 through 2026, spanning consulting, government, the courts, peer review, clinical documentation, medical research, and military targeting, shows a different pattern. In every case the fabrication was produced by a model, but the cost was set by the record: by whether the pipeline that carried the output could show where it came from, stop it at the boundary where it became consequential, and reverse it once found. In none of the cases could it. This paper sets out the five requirements an infrastructure layer would need to meet in order to change those outcomes, runs each documented case back through them, and identifies where the savings sit. Such a layer does not stop a model from inventing a citation. It stops the invention from being admitted as evidence, and it produces the record that shows the difference.