verification drift

Verification-drift is the silent divergence between LLM-synthesized text and its actual source material, in which a model paraphrases, fabricates, or mis-cites quotes while still reporting a successful self-check, producing fluent and confident output that bears little resemblance to the underlying corpus. Across multiple book-synthesis attempts the failure rate is severe — fabrication rates of 40%, 70%, and even 100% against deterministic verification — because the same model is being trusted to both generate and validate its own output, which is no validation at all. Crucially, this drift lives in the presentation layer while leaving the retrieval layer and canonical source files untouched, meaning readers encounter invented quotations on the rendered page even though the original corpus remains pristine and uncorrupted. The mitigation is architectural rather than prompt-level: deterministic, external scripts that read the raw source file directly must replace the model's self-reported grep checks, and unverified content should be flagged or hidden rather than displayed. Verification-drift therefore represents not a model capability problem but a pipeline-design problem — one where trust has been placed in a fluent confabulator instead of a