Silent Degradation in Document Processing Pipelines: Two Independent Case Studies

Abstract

A synchronisation checker reported complete success while 45% of the artefacts it verified had lost their substantive content. The checker was correct: it compared what a generator had written against the source, and those fields matched. It had no occasion to ask whether they were the whole of what should have been written. We describe a class of failure in document processing pipelines characterised by three properties in combination — the output is plausible, verification passes, and the defect is detectable only against the source — and show instances of it in two systems sharing no code, designer, language or year of construction. One is a closed commercial pipeline producing legal opinions; the other a public tool for document integrity verification, in which three such failures arose during a single day of construction and are reproducible from a repository. None of the ten defects required an adversary; most arose from libraries behaving as documented. We argue that verification is structurally unable to detect this class, since it establishes consistency with an expectation rather than with a source, and that introducing language models removes a signal a pipeline formerly had: a stage that once failed on degraded input now answers plausibly instead.