The document had nine sections, forty-one tables, and not one fact in it.
Section one, technical analysis: innovation, maturity, security assumptions, performance โ four rows, four N/As. Section two, token economics: team allocation, early investors, community float, treasury โ blank, with a parenthetical noting that Ponzi-structure risk could not be judged and that no charitable assumption would be made in the absence of data. Sections three through nine repeated the pattern. Market positioning, N/A. Ecosystem dependencies, N/A. All four prongs of the Howey test, N/A, with a composite verdict reading "cannot determine." Team capability, N/A. A six-category risk matrix with six empty rows. At the bottom, an information-value rating of zero stars out of five, awarded four times over.
One line in the middle of it was worth more than the rest of the document combined: in a due-diligence context, missing information should be scored as the highest-grade risk โ information asymmetry, unverifiability โ not treated as a neutral default.
That sentence is correct, and almost nobody applies it. What I was holding was not a research report. It was a refusal, formatted to look like a report, produced by a pipeline that had correctly detected it had been handed nothing and stopped.
The uncomfortable part is not that it stopped. It is how rare that behavior has become.
I have been on the other side of this document for nine years. In 2017, at twenty-four, I was a junior analyst in Melbourne reading whitepapers โ fifty-one of them, by my count, in a single eighteen-month stretch. The bottleneck then was human attention, and human attention is a natural rate limiter. You cannot skim fifty-one tokenomic models in a quarter and walk away with clean convictions. The friction was the control. Bitconnect taught me what happens when you remove it: I spent four months reverse-engineering failed allocation schedules and concluded that technology without a legal and financial substrate is not an investment, it is a lottery ticket with a whitepaper stapled to it.
By 2020 the job had changed shape. DeFi Summer turned research quantitative. I was modeling yield strategies for Aave and Compound, then retreated into a corner to study impermanent loss in ETH/DAI pools and wrote up liquidity fragility in Uniswap V2 โ depth, slippage, the specific way leverage disguises itself as depth. The lesson from that period never left me: yield is usually risk wearing a different coat.
Then the ETF approvals in 2024 handed research a new customer. Institutional allocators do not want narrative. They want coverage โ standardized, comparable, repeatable, delivered on a schedule, in a format a compliance committee can file. That demand produced the nine-dimension template: technical, tokenomic, market, ecosystem, regulatory, team and governance, risk, narrative, supply-chain transmission. Score each, weight each, publish. The template is a genuine improvement over vibes, and it is also a compression of judgment into checkboxes โ which means whoever automates the checkboxes automates the compression.
Which brings us to the present. Every serious desk now runs some layer of machine assistance over that template, and the economics of that layer are badly skewed, because of two failure modes the empty-input document named explicitly.
The first is hallucinated analysis: confident, well-formatted, structurally plausible conclusions generated in the absence of any factual anchor. In a research context this is not an annoyance. It is a financial hazard, because at the point of consumption it is indistinguishable from real analysis.
The second is the absence of fail-fast behavior. In systems design, fail-fast means that when an unrecoverable input error is detected, the process terminates immediately and loudly rather than continuing with a corrupted state. The empty-input document was a fail-fast. Almost no production research pipeline is one, and the reason is not technical.
Start with the template, because the template is where the deception lives.
A due-diligence framework is a control surface. Each field is a checkpoint where an analyst either clears an item or flags an exception. In a financial audit, a blank workpaper field is an exception by definition โ it cannot be signed off, it cannot roll into the opinion, it must be cleared or disclosed. The blank is the loudest thing in the file.
Crypto research inverted this. Here a blank becomes a row, the row becomes a table, the table becomes a section, the section becomes a document with a header and a star rating โ and the star rating is the only part anyone reads. The template reproduces the shape of rigor without the substance of it. A compliance officer skimming a Howey table sees that a Howey analysis was performed. She does not see that all four prongs returned N/A and the composite verdict was "cannot determine," because that verdict is rendered in the same font as a real one. The performance of diligence becomes the product, and the performance is cheap.
I watched this mechanism work from inside it in 2022. When the lending market came apart, I spent three months in near-total isolation auditing the balance sheets of three major lending protocols. The stated collateral was diversified across a dozen assets and four custodial arrangements. The actual exposure was correlated โ the diversification existed on the label and not in the correlation matrix. The number that mattered was not on any dashboard the protocol published. It was in a footnote about who held what, and the footnote had been read by approximately nobody, because the dashboard had a green number on it and the footnote had a grey one.
That is the mechanism stated precisely. A missing number and a fabricated number look identical in a table, and only one of them costs money to produce.
Which is why the model fills the gap. Language models are trained toward fluent completion, and human feedback rewards helpfulness. "Insufficient data" is a low-probability sequence in that distribution. An analyst who says I cannot assess this is not billable; a system that returns nine fully populated dimensions is. So the default behavior of an automated research layer is fabrication, and the cost structure is brutally asymmetric: producing a plausible TVL figure takes milliseconds, and disproving one takes an analyst roughly a day and a half โ assuming she catches it at all. By the time she does, the figure has usually been laundered through a deck, a memo, an LP call, and a governance forum post, and each hop strips another layer of provenance until the number has no author at all.
Token economics is where this hurts most, because unlocks are the one dimension that is forward-looking and therefore unfalsifiable at the moment of decision. A fabricated twelve-month cliff and an actual thirty-six-month cliff produce identical charts in month one and catastrophic differences in month fourteen. Nobody is lying at the point of sale. They are simply reciting a schedule that was generated to fill a cell.
The laundering chain matters most when it terminates in a governance vote. A DAO with the legal status of a group chat puts a treasury allocation to a token vote, the supporting analysis contains three fabricated figures and one hallucinated competitor comparison, quorum is met by eleven wallets, the transfer executes. When it goes wrong there is no legal person to sue โ there are token holders with unlimited personal exposure and a snapshot of a decision that was, in the strict sense, made by nothing. No legal status upstream plus no provenance downstream is not a governance model. It is an unpriced liability with a Discord server.
The same pathology runs through infrastructure. A rollup can publish a dashboard showing rising TVL and rising transaction count while its operator's actual margin is deeply negative on proving costs, because proving costs do not appear in any metric the dashboard was designed to display. The dashboard measures what is easy, not what is true. And it will keep measuring that until gas returns to levels that make the arithmetic survivable, which is not a technical question and will not be solved by a better proving system.
Bitcoin has the equivalent problem in a different register. Post-ETF, the marginal price-setter is an instrument with creation and redemption mechanics that most published research does not model at all. I have read correlation studies asserting a clean decoupling between ETF flows and global M2 on samples short enough to be noise, and I have watched those studies get cited as structural findings within a week. The research was not wrong because the author was careless. It was wrong because the pipeline required a finding and the sample offered a coincidence.
So what does a real control look like? Not a better model. A minimum input set, enforced before inference ever runs. Five fields, and they are not bureaucracy โ each determines which of the nine dimensions can be weighted and which must be skipped entirely.
At least three sourced information points, each traceable to an origin. Project or protocol identity, so competitive and regulatory lookups can be anchored rather than guessed. Title and source media, because an official blog post, a second-tier outlet, and a pseudonymous account on a social feed carry radically different priors and must be weighted differently. A publication timestamp, because the most common analytical error in this industry is not a wrong number โ it is a right number treated as new when it is three weeks old and already priced. And an article type: technical release, funding announcement, governance proposal, opinion, or price commentary. Those five categories map onto entirely different weights, and conflating them is how a funding round gets read as a product.
Supply those five and the pipeline can work. Fail to supply them and it must refuse โ loudly, in a document that says so, with the null preserved rather than papered over.
Here is the conclusion I have actually come to believe, and it comes out of the work I did this year on decentralized compute markets. The missing primitive in both AI training data and crypto research inputs is provenance, not intelligence. We are spending enormous capital building permissionless compute networks to feed models โ interviewing developers, mapping data pipelines, arguing about sovereignty in training corpora โ while the research those models emit carries no chain of custody whatsoever. We demand validity proofs from Layer 2 operators and accept vibes-with-tables from the analytical layer that decides where capital goes. The asymmetry is absurd on its face. A rollup that cannot produce a proof gets delisted. An analyst note that cannot produce a source gets published, syndicated, and cited.
The consensus fix is more intelligence: larger context windows, better retrieval, finer tuning, richer datasets. That is solving at the wrong layer. The empty-input problem is not a capability failure. It is an incentive failure, and it will survive every model upgrade.
Follow the money. Research is paid for coverage, not for accuracy. A desk covering forty tokens earns more than a desk covering twelve and abstaining on twenty-eight, regardless of which desk was right. Editorial metrics reward information gain โ a term I use deliberately, because it has been quietly converted from an epistemic standard into a content quota. Every incentive in the stack pushes toward confident novelty and away from the only honest sentence available: we do not have enough to assess this.
Which produces the genuinely counter-intuitive conclusion. The all-N/A document was the most accurate research artifact in the entire pipeline, and the pipeline flagged it as a system error. We have built an industry in which the only truthful product is classified as a bug.
And the blind spot is not that our models hallucinate โ everyone knows that now. The blind spot is that we have no auditor for the inference layer itself. We audit code, treasuries, validator sets, proving systems. Nobody audits the chain of reasoning that converts an on-chain observation into an allocation decision, and that chain is precisely where a fabricated figure becomes a position. The next decade's systemic risk will not originate inside a smart contract. It will originate in a research stack that was never designed to say no.
The differentiator for the next cycle will not be who has the smartest model. It will be who can afford to abstain โ who can sell a null result as a product, price the cost of finding out, and hand a compliance committee a document that says: we do not know, here is what it would cost to know, and here is what we are not buying until we do.
Ask a simpler question of your own stack. If it cannot produce a null, what exactly is it producing?
Emotion is the asset; discipline is the hedge. But a hedge you cannot verify is not a hedge โ it is a story you told yourself in a bull market.