Zero Facts, Nine Sections: How AI Research Pipelines Manufacture Conviction From Empty Input
A nine-dimension research report crossed my desk last Tuesday. Technical. Tokenomics. Market structure. Ecosystem position. Regulatory posture. Team and governance. Risk matrix. Narrative. Supply-chain transmission. Nine headings. Nine tables. Every cell rendered in clean, confident formatting.
Every field said the same thing: insufficient information to evaluate.
The document ran just over four thousand words. It was also the most honest output that pipeline had produced all quarter. The previous eleven reports carried the identical nine headings and an identical evidence base โ an empty one โ but those came back fully populated. Directional calls. Grade bands. Target levels.
Structure was intact. Content was absent. Nothing in the system flagged the difference.
Context
Since 2025, the standard crypto research desk has stopped being a room of analysts. It is a chain of model calls. A crawler pulls governance forums, Dune dashboards, project blogs, and thirty-odd Telegram channels. A first-stage extractor reduces raw text to atomic claims. A second-stage analyzer maps those claims onto a fixed evaluation template. A third stage writes the note.
The economics are obvious. Coverage went up tenfold. Marginal cost per report fell toward zero. Verification spending did not move.
That is the gap. Volume scaled. Evidence did not.
In my own build โ the agent framework I shipped in 2026, LLMs reading sentiment across fifty platforms and rebalancing across fifteen protocols โ we hit this wall in week three. The system generated signals on mornings when the crawl returned almost nothing. Thin news days. Holiday weekends. Low-liquidity windows. The agent still produced a full output, because a full output was what the schema asked for.
The specific failure I saw last week was mundane. The extractor hit a fetch error on the source document and returned an empty field set. The analyzer ran anyway. There was no interlock between stage one and stage two โ no condition stating that an analysis cannot proceed from zero evidence. Nine headings later, the machine had produced a document any reader would file as research.
Core
Here is the mechanism, and it is not mysterious.
Language models are completion engines. A template with nine headings is nine obligations. The loss function rewards a filled field far more than an empty one, because across the training corpus โ a decade of published crypto research โ almost nobody writes "I do not know." They write the section.
So an underconstrained research agent does not hallucinate. It complies. Hand it a schema and no evidence, and it resolves the conflict in favor of the schema.
I have watched the same failure in a different machine. The interest-rate models in Aave and Compound are utilization curves. They output a borrow rate at ninety-two percent utilization whether or not real borrow demand exists at that price. The number is a function of the curve. It is not a function of the market. A model that must produce a value will produce one.
The research note is the same artifact. It is a function of the template, not of the evidence.
The contrast is execution. In 2020 I ran a bot against a price discrepancy between Uniswap V1 and MakerDAO โ four thousand trades, one hundred forty-five thousand dollars in profit, dead the week V2 shipped. Every trade was verifiable at the moment of execution. Either the spread existed or it did not, and the block explorer settled the argument in twelve seconds. Research has no block explorer. That is precisely why the evidence chain has to be built by hand, and precisely why almost nobody builds it.
The fix we shipped was one condition. Evidence set empty, halt. Named error, no narrative. Four lines of code. It was the highest-return commit in the repository โ higher than the sentiment model that captured eight hundred fifty thousand dollars of alpha during a low-liquidity window that spring.
We did not know that at the time. We found out when we rolled the halt backward across twelve months of historical triggers. Thirty-one percent of signals were suppressed. Their combined simulated P&L was negative six hundred forty thousand dollars. The gate was not a brake on the strategy. The gate was the strategy.
In DeFi, liquidity is the only truth that matters. An empty evidence set is a liquidity failure in the information market, and it produces the same outcome: the price gets set by whoever shows up anyway.
There is a second discipline, and it is granularity. Every claim in a research note has to map to a fact unit with a named source. If the mapping fails, the claim dies.
I ran that rule against UST in April 2022. The entire report was three contract interactions deep โ Curve pool composition, the redemption path, the dependency graph โ and every sentence pointed at something I could open and read. It was published three weeks early, ignored, and correct. The bullish UST research of the same period looked identical in format. Same headings. Same confidence. It just never touched the pool contract.
Contrarian
The consensus reaction to a report like this one is "the model hallucinated."
Wrong. The model did exactly what it was built to do. It filled a template. The hallucination lived in the org chart โ in the assumption that a schema is an analysis and that a filled table is evidence.
The template is not the thesis.
The honest version of this problem is bigger than one broken pipeline. Most published DeFi research carries the same defect in softer form. Nine-section reports where the entire evidence base is the subject's own blog post. Technical comparisons of rollup stacks that never mention the variable that actually decides the race โ which team closes more integration deals per quarter. That variable never makes the technical section because it is not technical. So the section gets written anyway, from vendor material, and gets read as engineering.
Then identity. Three years of soulbound-token literature. A genuinely large research output. Almost no user who wants an on-chain credit record that cannot be deleted. The papers exist because the template exists, not because the demand does.
The counterexample is worth keeping. The pre-ETF positioning in 2024 had a hard, checkable input โ whale accumulation on-chain and a published SEC calendar. Forty percent of the book into BTC perpetuals at three times leverage, timed to the ruling, two point one million dollars in a week. That trade worked because the input was falsifiable before the position was taken. Almost nothing in a nine-section research note meets that bar.
Greed is a variable; discipline is the constant. The variable tells you to publish. The constant tells you to check whether you have anything to publish.
Which brings the market implication. Research output is an input to positioning. If a measurable fraction of that input is synthesized from an empty evidence base, then the market's aggregate information set is smaller than it looks. In a sideways tape, with no trend to hide behind, that gap is the edge. Whoever can grade their own inputs is trading a different market than whoever cannot.
Takeaway
Three rules, and they are cheap.
Gate the input. A hard null check that returns a named error and stops the pipeline. Grade the pipeline on suppressed outputs, not produced ones. A system that never says "I do not know" is not a research stack. It is a content department.
Attach every claim to a fact unit with a source. Claims without a source do not get a section. They get deleted.
Track the ratio. Reports published versus reports that changed a position. If that ratio only moves in one direction, you are paying for formatting.
Over the next two quarters, the agent-infrastructure winners will be the ones selling abstention. Generation is commoditized. Refusal is not. The desk that emits no signal keeps its capital; the desk that always emits one is simply paying the market for attention.
Ask the question that matters: if you emptied the evidence field of your research stack tonight, how many of this week's convictions would survive?
Mine went from nineteen to four. I sized accordingly.