At 06:42 on a Tuesday in Tallinn, a research report landed in my inbox with all nine sections filled in. Technical. Tokenomics. Market. Ecosystem positioning. Compliance. Team and governance. Risk. Narrative. Supply-chain transmission. Every heading present. Every subheading formatted. Every table bordered in that particular shade of grey that only automated document engines produce.
Every cell said N/A.
Reading the room in a room of code, I did what I always do: I opened a terminal and checked whether the emptiness was real. Twelve lines of Python — a parser that walked the heading tree, stripped the markdown, and counted tokens that were not "insufficient information." The script came back with 41 non-null tokens across roughly 4,100 words, and thirty-one of those were the disclaimer at the bottom.
The document had been produced by a two-stage analysis pipeline. Stage one was supposed to extract facts: title, source, claims, protocols referenced, timestamps. Stage one returned nothing. Stage two then produced a cathedral on that nothing, complete with load-bearing columns and a floor plan.
Nothing in the cathedral announced itself as fiction. That is the part worth your attention.
The Industrial Demand For Coverage
For most of the last decade, crypto research was an artisanal trade. You read the whitepaper, you read the contracts, you wrote a thread, you argued with strangers. The 2021 cycle ran on that. I ran on that — three parallel threads on Bored Ape, CryptoPunks and Art Blocks, weeks of interviews with collectors, an argument that the real product was digital identity rather than art. When floor prices collapsed in 2022, the identity thesis survived. Prices are a poor measure of narrative value, and that distinction is the single most useful thing that cycle taught me.
Then the institutional door opened. Spot Bitcoin ETFs. Custody desks. Compliance teams who needed coverage on assets their risk committees had never heard of. Between 2024 and 2026 the number of tracked digital assets roughly doubled while the number of people paid to actually understand them did not. I lived inside that gap at a Tallinn consultancy, translating on-chain behaviour for analysts who think in earnings multiples. My report on long-term holder behaviour — the piece arguing dormant coins were quietly functioning as collateral in stablecoin markets — got cited by three traditional finance firms. Not because the finding was revolutionary. Because it arrived in a format their internal processes could digest.
That is the lesson I keep returning to. Institutions do not read claims. They read scaffolding. Headings, tables, confidence labels, disclaimers. The container does most of the persuading, and the container is trivial to fabricate. Which means the cheapest way to satisfy institutional demand for coverage is not to produce knowledge. It is to produce the appearance of knowledge at scale. Every model released since 2023 has made that cheaper.
Stage one goes empty. Stage two fills the frame. Nobody notices, because the frame is exactly what the buyer was told to expect.
What A Template Does When Its Subject Disappears
There is a technical way to describe what happened, and it borrows from the cryptography I spent my undergraduate nights on at Tartu.
A zero-knowledge proof has two properties that matter. Completeness: a true statement can be proven. Soundness: a false statement cannot. In 2020 I was writing Python verifiers for Zcash's constructions, mostly badly, occasionally successfully, and what stuck with me was how much of the security lives in soundness — in the machinery that refuses to accept.
A nine-dimension analysis template is a proof system. Each dimension is a question that presupposes a subject: which consensus mechanism, what unlock schedule, who are the lead investors, which jurisdiction. When the subject exists, the template is a completeness engine. It forces the analyst to answer all nine, which is genuinely useful. When the subject is missing, the template does not halt. It answers with the shape of an answer. The technical dimension returns "not assessable," the tokenomics dimension returns "not assessable," and the document ships with nine completed sections describing a void.
The template has no soundness clause. Nothing in its architecture says: if the input is empty, the output is invalid. The failure mode is not error. The failure mode is fluency.
I have seen this before in a different costume. In 2022 I spent six months building a mental model of modular architecture, reading data availability sampling research and publishing illustrated explainers until my Substack reached five thousand readers. The most useful thing that exercise gave me was a permanent allergy to availability rhetoric. Most rollups I have looked at post fewer bytes of actual data per day than a mid-sized bank's internal messaging system, while their documentation reads as though they invented bandwidth. The format of data availability is not data availability. It never was.
How A Null Becomes A Number
The empty report itself is a paperweight. The catastrophe is the path it takes downstream.
Trace it. An automated pipeline generates the document. A summarisation layer compresses it into a dashboard card. The dashboard card feeds a portfolio monitoring tool. The monitoring tool feeds a model a trading desk uses to size positions. And since 2025, increasingly, the monitoring tool feeds autonomous agents that execute without a human in the loop.
Now consider what a language model sees when it reads the line "Risk level: N/A — insufficient information." To a careful human reader, that is a refusal. To a pattern-matching ingestion layer optimising for structured extraction, "N/A" is a value in a risk field. And a risk field populated with something other than "high" reads as, at minimum, unremarkable. Absence launders itself into neutrality.
I spent part of 2026 auditing agent-based trading frameworks — the technical half of the autonomous-economy narrative I have been writing about. Fourteen frameworks. Nine of them, when I fed them deliberately incomplete market data, treated missing fields as bearish-neutral rather than unknown. Two treated a missing field as a zero. One propagated the null straight into a position-sizing calculation and opened a trade.
None of them logged that they had been flying blind. That is the contamination vector. It is not that the model hallucinates a fact. It is that the model hallucinates the absence of a problem.
A null result that is never flagged is indistinguishable, at the point of decision, from a clean bill of health.
Provenance Is The Whole Product
Go back to the metadata that stage one was supposed to extract: title, source, publication time, protocols referenced, claims with data attached. That list is not bureaucratic hygiene. It is the entire epistemic content of a report.
A claim without provenance cannot be priced. You cannot weigh it, discount it, or update against it. You can only repeat it. And a pipeline that loses provenance while preserving format has produced something structurally identical to a rumour — a rumour with excellent typography.
I keep landing on the same observation about governance. Last quarter I traced a proposal through a DAO governing roughly $400 million. Turnout was 3.1%. Two wallets carried 41% of the quorum. The forum thread was substantive, the proposal page was immaculate, execution was automatic, and the outcome was decided by an electorate smaller than a mid-sized Discord server. The format of decentralised decision-making was fully intact. Participation was a rounding error.
The empty report and the empty quorum are the same object. Both are containers that kept operating after their contents left the building. Both were trusted because they looked the way trust is supposed to look.
The Fix Is A Halt Condition
The instinct when a pipeline hallucinates is to improve the model: better prompts, better retrieval, a stricter evaluator. That is the wrong layer.
The right layer is an assertion. If extracted information points equal zero, the pipeline halts and returns a null result with a reason code. Fail closed, not open. The cost of halting is a delayed report and a slightly embarrassed operator. The cost of not halting is a position sized on a document that describes nothing, held by a desk that believes it read something.
I have started asking a different question of every research product that crosses my desk. Not "is this analysis correct." Rather: what share of this document is stage-two-over-empty — structure generated to fill a frame whose subject never arrived? Framed that way, a lot of what circulates looks different. The confident market note with no on-chain evidence. The thread with a thesis chart and no methodology. My own early work, if I am honest, in the weeks before I learned to run the numbers myself.
The Uncomfortable Counterargument
Here is the part that cuts against everything I just argued.
The empty report is safer than the full one.
A padded report — 20% fact, 80% inference, rendered in the same confident register — is indistinguishable from analysis without doing the work. It has numbers. Numbers create the illusion of resolution. A reader cannot easily tell where the evidence stops and the narration begins, and neither, frequently, can the person who wrote it. That document enters the decision chain wearing a suit, and it takes a hostile reader with a terminal to undress it.
The empty report at least announces itself. Every cell says N/A. It is a smoke detector that fires before the fire.
Which means the market has the incentive gradient pointing precisely the wrong way. A report that says "the input was empty, we cannot analyse this" earns zero engagement. A report that invents a twelve-dollar price target and dresses it in three technical indicators earns ten thousand impressions and a podcast invite. The reward is for coverage, not accuracy. Under that gradient, honesty is a cost centre — unless the reverse becomes true at the institutional layer, where the buyer is a risk committee rather than a timeline.
I don't think AI analysts will replace human ones. The replacement question is a distraction. The scarce capability, in a market where every participant is paid to have an opinion, is the ability to say "I don't know" — and to have that refusal survive the pipeline, the summariser, and the agent. That is a systems property, not a personality trait. It has to be engineered.
What I Expect To Be Measured By 2027
Within eighteen months, I expect at least one serious research platform to publish a null rate alongside its coverage count: the share of documents that terminated with no conclusion because the underlying evidence was insufficient. It will look like an audit report. It will be read like a credit rating.
The analysts who survive the next compression of this industry will not be the ones who can write nine sections about anything. They will be the ones whose pipelines stop. Reading the room in a room of code, the most valuable output is not the answer. It is the circuit breaker that refuses to give one.
How much of what you read this week was stage-two-over-empty — and did anything in the document tell you?