In a long brief or a multi-step agent, the model builds on what it already wrote. An invented case introduced in the statement of facts gets cited again in the argument and relied on in the conclusion — a single fabrication compounding into a structural one.
Probability map. Many people are under the mistaken impression that an LLM is some sort of hypercomplex thinking machine. Really it’s just an incredibly huge probability map of words, patterns, and phrases followed by other words, patterns, and phrases. Unless it’s attached to a database of authoritative sources, nothing connects what it generates back to reality.
Completion pressure. Every model runs under a standing directive that strongly favors providing an answer rather than tell you it can’t. Hand it a gap where a real authority should be, and it fills the gap instead of flagging it.
Latent contradiction. Your LLM prompt can look fine on its face, but against the backdrop of law, non-obvious contradictions can be hiding in wait. That conflict is invisible to you, and sometimes even to the model itself. Something has to give, and completion pressure decides what: it honors the request and manufactures the rest.
Long-form generation is self-referential. As a model writes a thirty-page brief, its earlier output becomes context for its later output. A case it invented on page four is now, as far as the model is concerned, an established part of the record — so it cites it again, characterizes it further, and leans on it in the argument.
Agentic workflows amplify this. When a model plans, drafts, and revises across steps, each step trusts the last. A fabrication introduced early isn’t re-examined; it’s inherited. By the end, the invented authority isn’t a stray cite — it’s woven through the structure of the argument.
That’s why longer, more autonomous generations carry more risk than short ones. The problem isn’t only that there are more cites; it’s that early errors get reinforced instead of caught, and the model’s confidence in its earlier work rises with every reference back to it.
The same cause, a few ways it turns up in ordinary practice. Each is routine work you’d never flag as risky — which is exactly how the cause slips in unnoticed.
The shape of any long document — a case named once in the background, a premise stated up front, an assumption the argument keeps leaning on twenty pages later.
By the argument section it’s cited twice more and treated as settled authority — one fabrication, three appearances.
The default of agentic drafting — a plan-then-write pipeline, a draft-then-polish loop, any workflow where each step builds on the last without re-checking it.
The refinement polishes the prose around a fabricated cite instead of questioning it; the error survives every pass.
The way a summary grows — a characterization introduced early, restated in the roadmap, expanded in the argument, each pass drifting further from the source.
Each later reference drifts further from any real source, all tracing back to a passage that was never real.
Adding a citation directive (“only cite real cases”) mathematically cannot ever fix the problem. Humans read a citation directive and think the LLM must comply. But every LLM has completion pressure built into its architecture, a mathematical property of the model itself, that will readily override such a directive.
Verbatim checks the finished document as a whole, so a fabricated authority that appears three times is caught all three times, not trusted because it recurs. Repetition inside a draft is not evidence — the report weighs every citation against its source independently, no matter how load-bearing the argument has made it.
For workflows that draft in steps, the same engine runs on the output before it reaches a person, so an error introduced early doesn’t ride the whole chain into a filing. The check happens once the draft is whole, where the compounding is finally visible.
Verbatim reads a finished brief and reports, for every authority it cites, whether the cite is real and whether the quoted language actually appears at the pin cite — so a fabrication surfaces on your screen, not in a show-cause order. Bring a brief and we’ll walk you through the report.