Degradation in Transit: Zero Hallucinated and Six Corrupted Citations in Three LLM-Condensed Manuscripts
Abstract
Across 70 references in three manuscripts condensed from our own longer papers by a language model, no reference was fabricated and six were corrupted: real works whose entries no longer describe them accurately. A dropped co-author. A forename changed from Jaime to Thomas. An author list cut from eight names to three. A year that contradicts the volume number printed beside it. Every one of the six cites a work that exists, and every one resolves, so a checker that verifies a DOI or an arXiv identifier passes all six. Two entries appear in two of the three manuscripts, condensed separately from the same source bibliography; each is correct in one manuscript and corrupted in the other, which makes the corruption per-transcription and stochastic and means a second condensation pass is no remedy. All six faults were introduced by the condensation, and none was inherited from the source bibliography. This is one condensation pipeline, one author group, 70 references and no control condition, so we report counts in a corpus and no estimate of a population rate. The measurable failure mode here is degradation in transit. Nothing was fabricated at source, and existence verification is blind to what did go wrong.