Last week I reviewed a due-diligence report. Forty-one pages. Nine analytical dimensions. Six comparison tables, a risk matrix, a confidence assessment, a disclaimer, and a closing section titled 'Guidance for Resubmission.' It was generated from an empty input. Not thin. Not paywalled. Not redacted. Empty โ no project name, no token symbol, no source, no claim, not one sentence of original text.
The pipeline did not crash. It produced.
Every table had rows. Every row said the same thing in a different font: N/A โ information insufficient. The supply schedule was blank. The Howey test was blank. The competitive landscape was blank. And at the top of each of the nine sections, a confidence label: High. Because the system was, in fact, highly confident that it had nothing.
I have spent ten years watching crypto manufacture conviction out of air. I had not, until last week, watched a research stack do it with correct grammar and a compliance footer.
That is the interesting part. Not the bug. The output.
Context
The institutionalization of this market changed what a research process is supposed to look like. Before the spot ETF approvals, diligence in crypto was a Discord channel, three whale wallets, and a spreadsheet. After them, allocators needed something that survived an investment committee: standardized dimensions, documented sourcing, reproducible scoring. Frameworks proliferated โ nine-dimension token screens, twelve-point protocol checklists, red-flag matrices with weighted scores.
Then the LLMs arrived and ate the middle of the workflow.
The architecture most funds now run is a two-stage pipeline. Stage one extracts: it reads a source document and emits structured facts โ an information-point array, a project identifier, a thesis summary, a domain tag, a timestamp, a source-quality judgment. Stage two analyzes: it consumes that array and produces the nine dimensions.
I have built versions of this myself. In 2024 I led three analysts through a fifty-page report on Bitcoin ETF flows, tracking roughly $2 billion of net inflows in the first month and correlating them against S&P 500 volatility surfaces. The extraction layer was the part that broke constantly. Not the modeling. The extraction.
Which is exactly where this report failed. Stage one emitted an empty array. Stage two received it and, instead of halting, rendered the void as a deliverable.
The output is structurally honest and operationally toxic, and those two properties are not in tension โ they are the same property viewed from different desks.
Core
Here is the mechanism, and it matters more than the anecdote.
A pipeline like this validates on schema, not semantics. The information-point field is typed as an array. An array of length zero is a valid array. The project-name field is typed as a string; 'Not provided' is a valid string. Confidence is typed as an enum; HIGH is a valid enum value. Every field passes. The job returns green.
So the report is not a hallucination. That is what makes it dangerous. A hallucination is detectable โ it makes a claim, and claims can be checked. This document makes no claims at all. It is a schema-valid null, and nulls do not trigger alarms because nothing in the validation layer is looking for absence. Systems are built to catch wrong answers. Almost nothing is built to catch the confident non-answer.
Run the arithmetic. Suppose a given analytical dimension has, say, a 15% chance of producing an unsupported assertion when the underlying facts are absent โ the model fills a gap with a plausible competitor, a plausible unlock schedule, a plausible team background. Nine dimensions. The probability that at least one fabricated claim survives into the final PDF is one minus 0.85 to the ninth power, which is roughly 77%.
The empty input is the highest-risk input in the system, not the lowest. Everyone assumes the danger scales with how much you feed the machine. It scales with how little you feed it and how much structure you demand anyway.
There is a second-order problem, and it is the one that will actually cost money. In compliance terms, 'not assessed' and 'assessed, no issue found' converge the moment they are printed in the same typeface under the same heading. A portfolio manager skimming at 11 p.m. sees a section header, a table, and a conclusion. She does not see the difference between a filled cell and an empty one unless someone forces her to. The document's presence is doing the persuading, not its content.
A risk matrix with no probabilities is a spreadsheet of adjectives. It reads as diligence. It functions as decoration.
I learned this the hard way, twice. In 2017 I spent three months hand-tracking whale wallets on Etherscan and logging suspicious token launches into a personal spreadsheet โ over fifty of them. Eighty percent died from tokenomics, not code. The lesson I took was about distribution, but the lesson I should have taken was about records: I had a spreadsheet full of rows, and for months I mistook the row count for evidence. In 2020 I put $5,000 across five DeFi protocols during the farming mania and lost 30% in a single flash crash. Same error, different wrapper. I had positions and I had dashboards, and I assumed the dashboards were the analysis.
Liquidity is a ghost, not a foundation. So is a completed template.
The engineering fix is trivial and almost nobody ships it. Enforce a minimum item count on the extraction array. If the count falls below three, and none of the entries carry a resolvable source, the pipeline must halt and return an error โ not a report. Refuse to render. Make the null loud. The cost of a circuit breaker on an empty array is one line of validation. The cost of the alternative is a document that will be filed, cited, and eventually discovered by someone who assumed it meant something.
Contrarian
The consensus worry about AI in research is fabrication โ the model inventing a partnership that never existed. That worry is real but overrated, because fabricated specifics are falsifiable and falsification is cheap.
The underrated failure is the confident null: a report that is technically accurate, fully compliant, and completely inert, produced at scale because the appearance of diligence is what the workflow actually rewards. Compliance does not require that you know something. It requires that you can demonstrate you looked. A forty-one-page document with nine dimensions and a High confidence label demonstrates that you looked.
This is where the industry's incentives are genuinely broken. A blank report is not a neutral artifact. It is an active claim โ it says the pipeline ran, the schema held, and no red flag emerged. Downstream, that reads as a clean bill of health, because nobody encodes the distinction between 'no findings' and 'no data' into the database.
Smart contracts don't create demand; they route it. Research pipelines don't create knowledge; they route whatever facts you feed them โ and when you feed them nothing, they route the shape of knowledge instead. That shape is the most tradeable lie in this market, because it is the one nobody thinks to short.
Takeaway
Sometime in the next eighteen months, a position will blow up and the post-mortem will trace back to a research pipeline that returned a complete, well-formatted, schema-valid report on an empty input โ and every reviewer will swear they read it.
The question is not whether your model hallucinates. The question is what your model does when the information-point array is empty. If the answer is that it writes the report anyway, you do not have a research process. You have a very expensive way of producing N/A.