In late 2025, a colleague forwarded me a research document. By every formal measure, it was complete. Eight analytical dimensions. A risk matrix with probability and impact columns. A regulatory test decomposed into its four prongs. Confidence markers. A disclaimer. Thousands of words of structured, competent prose.
At the top, in bold, sat a single admission: the input was empty.
Not redacted. Not corrupted. Empty. The first-stage parser had returned nothing — no headline, no source, no information points, no project name. The second-stage engine had responded by producing this cathedral of analysis anyway. Every section titled. Every cell filled. Every substantive value replaced by the same three characters: N/A.
There was even a section titled "Key Risks." The key risk, it turned out, was that there was nothing to analyze. The document identified this risk, named it, and then continued for another four thousand words.
I read it twice. The unsettling part was not that the pipeline had failed. The unsettling part was how professional the failure looked — how closely it resembled the research I am paid to read every week.
To understand why such a document exists, you have to understand the machine that produced it.
The industrialization of crypto research happened quietly. In 2017, when I was twenty-five and auditing the Golem whitepaper against its own reward-distribution math, "research" meant a human reading a document and modeling its incentives by hand. It was slow. It was lonely. And because it was expensive, it was scarce — which meant it was, on average, honest. Nobody spends three weeks modeling a token's fee volatility in order to fabricate a conclusion.
By 2020, the economics had inverted. Content volume became the metric. Protocols needed coverage; funds needed deal flow; exchanges needed listings justified; and an entire layer of "research" emerged whose product was not truth but legibility — the appearance of having looked closely.
The architecture that scaled this is now standard. A two-stage pipeline: first, a parser decomposes a source — an article, a filing, a governance post — into discrete "information points." Second, an analytical engine maps those points across a fixed grid of dimensions: technical, tokenomic, market, ecosystem, regulatory, team, risk, narrative.
The design is elegant. It is also fragile in a way that almost nobody checks. The framework assumes the parser succeeded. It has no branch for the case where the parser returns nothing. So when the input is empty, the second stage does what any well-trained language model does when handed a template: it fills the template. Not with facts — there are none — but with structure. And structure, it turns out, is enough to fool almost anyone.
A fixed-field template is not a neutral container. It is a force. Hand any generative system a grid with labeled cells, and the grid will demand to be filled. This is not a flaw specific to AI. It is a property of form itself. A form wants completion the way a low-pressure system wants air.
Consider what the empty-input document actually did. It could have returned a single line: "Insufficient input; analysis cannot proceed." That would have been the mathematically correct output. Instead it returned eight dimensions, each populated with the token "N/A" — and then, having satisfied the template's shape, it went further. It began to narrate its own emptiness. It explained that the first stage was empty. It listed what was missing. It recommended "the minimum input required to restore analysis."
That last move is the tell. The system could not simply stop. Faced with nothing, it generated a report about having nothing — which is still a report. The template won.
I have seen this exact gravity in tokenomics. In 2017, the Golem team published a reward-distribution mechanism that assumed transaction fees would remain stable. The assumption was never stated as an assumption; it was baked into a formula that looked complete. The formula had fields. The fields had values. Nobody asked what happened when the values were wrong. The template of "decentralized compute marketplace" demanded a reward curve, so there was a reward curve — and the curve's formal completeness concealed the fact that its central input, fee volatility, had never been measured at all.
That is the first failure mode: the template generates the output shape regardless of whether the input justifies it.
The second failure mode is linguistic, and it is subtler.
When the framework fills empty fields with "N/A," it collapses three entirely different meanings into one string. There is "N/A" meaning not applicable — the field genuinely does not apply. There is "N/A" meaning unknown — the data exists but was not retrieved. And there is "N/A" meaning unavailable — the analyst could not, or would not, determine it. These are not the same. One is a fact about the world. One is a fact about the pipeline. One is a fact about the analyst's restraint.
A human reading "N/A" in a risk matrix will assume the benign reading: that the category was considered and deemed irrelevant. The engine, meanwhile, may have meant something closer to: I have no idea, and I am not permitted to say so.
This is the quiet crime of automated research. It launders ignorance into neutrality. The reader sees a clean grid; the grid is actually a confession, and the confession has been formatted to look like diligence. Math does not care about your conviction, and it does not care about your template either — but a reader cannot tell the difference unless someone shows them.
I learned to watch for this in the DeFi Summer of 2020. When Compound and Aave were printing double-digit yields, the dashboards showed a number: APR. One number, one field, universally understood. What the field concealed was the difference between yield from real borrower demand and yield from subsidized token emissions — a distinction the single number could not represent and did not want to. I wrote "The Yield Trap" that year, arguing that one well-formatted percentage was hiding a systemic liquidity risk. The essay was unpopular. The liquidity crunch that followed was not.
The lesson repeats. A field that cannot distinguish "safe" from "unknown" will always be read as "safe."
The third failure mode is the most dangerous, because it is the most persuasive.
The empty-input document contained a risk matrix with columns for probability and impact. It contained a regulatory test broken into four prongs. It contained a five-star information-value rating. Every one of these is a form of rigor. None is a substance of it. A probability column with no probabilities is not analysis; it is the costume of analysis.
Behavioral economics has a name for why this works on us. We are pattern-completers. Shown a matrix, the eye seeks the missing cell. Shown a four-prong test with four blanks, the mind assumes a test was run. The apparatus of verification becomes a substitute for verification, because the apparatus is visible and the verification is invisible.
I watched this mechanism operate at institutional scale in 2022. Celsius published regular reserve attestations. BlockFi showed risk dashboards. Both wore the costume. Both had matrices. Neither had an invariant — a single number that could not be faked, like an on-chain balance, checked against the chain itself. When the costume came off, the absence underneath was total.
I spent three weeks in a cabin outside Austin that year, writing "The Illusion of Sovereignty," trying to explain why the form of decentralization had survived so long after the substance had left. The answer was the same then as it is now: we trust scaffolding because scaffolding is what we can see.
The fourth failure mode is the structural one, and the one I care about most.
In on-chain analysis, an empty input returns an empty output. If you query a contract for a balance and the contract does not exist, you get nothing. You do not get a beautifully formatted table of "N/A" balances. The chain is a ground truth, and ground truth does not decorate.
Narrative research has no such chain. Its reference point is the story itself. Which means that when the input is empty, there is nothing to contradict the output. The document cannot fail, because failure would require a standard, and the standard was never supplied. Narratives are liquid; truth is solid; and a pipeline that only handles liquid will drown the moment the solid is required.
This is why the empty-input problem is not really an AI problem. It is a verification-architecture problem. Any pipeline — human or machine — that lacks an invariant will, under pressure to produce, produce. The pressure does not have to be malicious. It only has to be structural. A deadline is structural. A template is structural. A business model that pays per document is structural.
I have spent eighteen years watching narratives survive their own refutation. The Layer 2 "decentralized sequencer" has been a slide in a deck for roughly two years now — a single operator node wearing the word "decentralized" like a coat. The word is the field; the node is the value; and almost nobody checks whether the two match, because the field is legible and the value requires work. The stablecoin story is the same shape, inverted: PayPal launched PYUSD not because it loved crypto but because it preferred to become a regulatory partner rather than a regulatory target. The narrative said "adoption." The invariant said "hedging." One of those is checkable. The other is a press release.
In the chaos, look for the invariant. If there isn't one, you are not analyzing a system. You are reading its marketing.
The pipeline is not a mistake. It is a rational response to an irrational demand.
By 2026, the research market had a volume problem. Search algorithms rewarded "information gain" — the promise of a new insight per article. Funds wanted coverage of every protocol. Exchanges wanted listings justified. Content became a throughput game, and throughput rewards the two-stage pipeline: parse cheaply, analyze cheaply, publish quickly. A single human analyst producing one rigorous brief per week cannot compete with a pipeline producing forty. The market did not choose rigor. It chose rate.
Under that pressure, the empty input is not an edge case. It is the expected case. Every source will sometimes fail to parse — a paywalled filing, a broken feed, a governance post with no text. A pipeline built for volume must handle those failures gracefully, and "gracefully," in practice, means "without stopping." The template was designed never to stop. That is its virtue in production and its catastrophe in truth.
I understand the temptation from the inside. There are weeks when the market gives you nothing — chop, sideways, no signal — and the incentive is to manufacture a signal from the absence. But chop is not empty. Chop is waiting. The problem is that waiting does not produce forty briefs. So the machine fills the void with structure, and the structure looks like work, and the work looks like insight, and somewhere in that chain the only question that matters — is there anything here? — stops being asked.
A pipeline that cannot ask that question will always answer it.
Now the turn, and it is uncomfortable.
The empty-input document is the most honest research artifact I have read in years.
It told me, at the top, that its input was empty. It did not pretend otherwise. It did not bury the null in a footnote. It said, in effect: I have nothing, and here is the shape of my nothing. Most of the confident analysis I read every week does not do this. Most of it has a real input — a headline, a funding round, a governance vote — and then quietly stretches that input across eight dimensions it cannot support. The empty document failed loudly. The ordinary document fails silently, and the silence is the product.
So the contrarian claim is this: the hallucination is not new, and it is not artificial. It is the default state of narrative finance. Humans hallucinated tokenomics in 2017, yield in 2020, sovereignty in 2022. The SEC's regulation-by-enforcement is itself a kind of empty-input document — a rule withheld and then enforced, leaving the market to fill the blank with whatever story it prefers. AI did not introduce the empty input. AI introduced a machine that will show you the empty input, if you build the framework to admit it.
The bug is not that the engine filled a template with N/A. The bug is that we built a template that cannot say "stop."
What does this mean for the reader waiting for direction in a sideways market?
Demand the invariant. Before you trust any analysis — human or machine — ask one question: what number here could be wrong, and who checks it? If the answer is "the template," you are holding a costume. If the answer is "the chain," you are holding evidence.
The next narrative will not arrive with a warning. It will arrive with a dashboard. Solitude is the price of clear vision; the crowd sees a moon, and I look for the model underneath. Quietly positioned while the world shouts — that has always been the only durable edge.