llm reliability

LLM reliability breaks down in predictable, recurring ways when models are asked to handle direct citation: they paraphrase or fabricate quotes during synthesis, and they cannot be trusted to verify their own output, since the same process that hallucinates a quote will hallucinate the confirmation that it exists in the source. Even high-sounding accuracy rates (75%, 83%) are misleading, because for attribution work the acceptable threshold is effectively 100% — a single fabricated quote attributed to a real author destroys credibility in a way that many correct ones cannot repair. The most useful architectural response is to separate the retrieval layer (which reads source text directly) from the presentation layer (which renders and paraphrases it), so that failures can be localized, degraded gracefully, and fixed at the right level rather than treated as a monolithic problem. Naming the pattern — synthesis-quote-verification-drift — matters because it converts anecdotal citation errors into a reusable category that tools, audits, and future transcripts can reference systematically. Taken together, the atoms argue that LLM reliability for citation is not a matter of better prompting but of structural