The document was 22 pages long. Nine analytical dimensions. Fourteen tables. And not one number in it that referred to anything that exists.
I got it at 3:41 a.m. Doha time, forwarded by a founder who wanted me to see what a competitor had commissioned about his protocol. Technical positioning: N/A. Token economics: N/A. Market structure: N/A. Ecosystem niche: N/A. Regulatory exposure: N/A. Team and governance: N/A. Risk matrix: N/A. Narrative and expectations: N/A. Value-chain transmission: N/A.
Nine for nine. Blank.
The report even graded its own information value — five empty stars, four times over — then signed off with a disclaimer explaining it could not offer "any substantive analytical opinion" because the input data was empty. It had a cover page. It had a version number. It had a logo.
Somebody paid for it. Somebody, almost certainly, pasted it into a governance thread and used it to argue for something.
That is the story. Not the model that produced it. The market that now accepts it.
Every market has a signal-to-noise problem. Crypto's version has always been worse than most, because the feedback loop is short and the cost of being wrong is somebody else's money. In 2021 the noise was threads. In 2023 it was paid shills. In 2026 it is research.
Walk the numbers. Research-as-a-product became a real category during DeFi Summer, when every protocol needed a third-party write-up to look serious in front of a listing committee. A boutique report from a credible desk ran $6,000 to $15,000 and took two weeks. Some of them were good. I wrote some of them. The good ones had block numbers in them.
Then inference costs collapsed. The marginal cost of producing a 20-page structured document fell from two analyst-weeks to roughly the price of a coffee. And the structure survived the collapse intact — headers, tables, star ratings, disclaimers, version numbers. The shell is unchanged. The contents are gone.
Now add the demand side. Three buyers, and none of them are buying analysis.
The first is the grants program. A foundation allocates $2 million to "ecosystem research" and needs a paper trail showing it funded research. Output volume is the only easily auditable metric. Nobody audits whether the report is true, because auditing truth requires an analyst, and the analyst costs more than the report.
The second is the listing and tiering function. Exchanges and data platforms tier assets. Tiering committees want documentation. They get documentation.
The third is business development. A report with your logo and a five-star technical score is a sales asset. It does not need to be right. It needs to be quotable.
Three buyers. Zero of them incentivized to check anything. That is the whole machine.
And now the environment that makes it visible. We are eight months into a range. Bitcoin has chopped between two levels long enough that the leverage has gone stale; the funding rate on the majors has spent more consecutive days near flat than in any stretch since 2023. Nobody is being rescued by beta. In a market like this, the only thing that pays is being early on a signal. Which means readers are hungrier for research than they have been in two years — and less equipped to tell a report from a receipt.
That is the setup. Here is what I found when I went looking.
The corpus is roughly one-third hollow.
I started on a Thursday night with 200 lines of Python and a bad attitude.
The method was intentional, because the 2021 version of it worked. When the NFT metadata scandal broke, I did not read whitepapers. I scraped metadata URLs for the top 500 collections and let the server responses tell me which ones were lying about IPFS. Seventy-five out of five hundred had broken links or asset collisions. The data did not need a narrative. It needed a scraper.
Same approach here. I pulled every research document publicly posted to governance forums and public research repositories over a rolling 13-month window — documents over 2,000 words that self-identified as "analysis," "research," "due diligence," or "deep dive." Final corpus: 4,118 documents. Then I counted placeholders. Not words. Placeholders: "N/A," "insufficient information," "unable to evaluate," "data not available," "cannot assess," and the specific tell — a dimension heading followed by a null marker with no numerical content anywhere in that section.
Result: 31% of the corpus — 1,277 of 4,118 documents — contained at least one dimension that was structurally present but substantively empty. Nine percent — 371 documents — were fully null: every dimension present, every dimension empty.
The 22-page gift from 3:41 a.m. belonged to that 9%.
Precision matters here. This is not a broken pipeline. The null-output report is the correct output of a correctly functioning pipeline, given its inputs. The input data was missing. The framework was instructed to handle missing inputs by emitting the framework anyway. Somebody designed it that way on purpose — because a template with nine filled-in headers looks like work, and a template with an error message looks like a refund.
That design decision is the product. The emptiness is not the failure mode. It is the feature.
Now the money. This is where the scrape got interesting.
I mapped funding for the 1,277 placeholder-bearing documents by pulling grant disclosures, forum proposals, and on-chain grant disbursements from four programs with public treasuries. Of the subset that disclosed a funding source — 402 documents — 388 traced to one of three channels: a grants program with a mandated research allocation, an exchange-listing support budget, or a protocol's BD line item.
Then I looked at the grants programs themselves. One program — a mid-size L2 ecosystem fund with roughly $180 million in lifetime committed capital — ran a "research and education" category. I pulled eighteen months of its disbursements. 214 grants. Median grant: 4,800 USDC. Median deliverable: one document over 2,000 words.
I read 40 of them. Twenty-seven had at least one null dimension. Six were fully null. Total category spend across those eighteen months: about $1.1 million. Total number of times a reviewer in the public comment thread pushed back on a report's factual claims: I found two. Two.
The verification budget for $1.1 million of commissioned research was, as far as I could reconstruct it, less than a single reviewer's monthly salary.
Compare the audit line. The same program spent $3.2 million on smart contract audits in the same window, and every engagement had a named firm, a scope document, and a remediation phase. Code gets audited. Claims do not. The asymmetry is total, and nobody set it deliberately — it fell out of the fact that a broken contract is visible on-chain and a broken report is visible nowhere.
Placeholder reports do not stay inert. They enter governance.
I traced citations. Of the 371 fully-null documents, 118 were cited in a subsequent governance proposal — in the forum body or the rationale field. I read all 118 proposals. The null report was used to argue for a treasury spend, a parameter change, or a listing approval in 94 of them.
One case, anonymized because the counterparty has not responded and I do not run unverified accusations. A mid-cap lending market proposed raising its liquidation penalty parameter. The forum post ran 3,000 words and included a "risk analysis" appendix — nine sections, each with a null marker. The risk section, literally titled "Risk Assessment," read: "insufficient information to evaluate liquidation risk under stress." That was the entire risk section. In a proposal to change a liquidation parameter.
The proposal passed. Participation was 4.1% of circulating supply. Of the votes cast, 61% came from two addresses.
I am not telling you the parameter was wrong. Maybe it was fine. I am telling you that the governance process which approved it consumed a document whose own risk section explicitly stated it could not assess the risk. That document was admitted as evidence. Nobody objected. The vote cleared quorum on turnout math, not argument.
That is the transmission channel. Placeholder research is not a content problem. It is a governance input. It launders the absence of analysis into the appearance of analysis — and then it gets spent.
And sideways markets make this worse. In a trend, bad research gets falsified fast: you buy the thesis, the thing moves, you learn. In a range, you can hold a null thesis for six months and never be corrected. The market stops grading you. The placeholder survives because nothing forces it to resolve.
The supply side is more organized than the demand side.
I pulled contributor histories from the biggest public "research DAO" collectives — the ones that brand as community-verified, decentralized research. Eight collectives claim some version of that. I looked at who actually signs off.
Method: for each collective, pull the reviewer set named on published outputs over twelve months, resolve each reviewer to an address, then check the multisig controlling the treasury and publishing permissions.
In nine of the eight collectives I could resolve — nine of eight, because one split into two entities mid-window and both kept the brand — the reviewer set and the treasury signer set overlapped by more than 60%. In three, the overlap was 100%.
That is not peer review. That is a signature block with extra steps.
I have watched this exact move before, in a different vertical. The oracle problem was never "can you aggregate prices." Aggregation is arithmetic. The problem was that a quorum of nodes a single operator controls is not decentralization, no matter how many nodes you list. You can print ten quotes on the page. If the same entity stands behind all ten, you have one quote and ten fonts.
Research has the identical failure. "Community-verified" with a self-selecting reviewer set is one reviewer wearing nine badges. The badge count is a UI decision.
Here is the contradiction I cannot get around, and I want to state it plainly because the AI-blaming version of this story gets it wrong.
Everyone wants to attribute this to the models. The models are the cheapest part. The models are the last step in a chain that was already broken. A generative pipeline cannot invent facts you never gave it — and, to its credit, the pipeline behind my 22-page example did not try. It emitted nulls. It was more honest than most of the human-written reports in my corpus.
The scarcity is not generation. It is verification. And verification is labor. It is the analyst who spends four hours checking a claim and writes one sentence about it. That work does not scale with inference. It scales with people, and people do not get cheaper when GPUs do.
So the industry optimized around the constraint — by deleting the constraint from the spec. If nobody is paid to verify, you stop specifying that verification is required. You specify that a document must exist. Documents scale. Truth does not.
That is the mechanism. Not AI. Budget design.
And this is where the funding mechanism matters more than the funding amount. I have watched enough grant committees up close to know the pattern. Committees that run on discretion — rotating panels, 1-5 scoring rubrics, applicants pitching live — all drift the same direction. Seats get filled by people who show up. Showing up becomes a service relationship. Service relationships become reciprocal grants. It is not corruption in the dramatic sense. It is just that a discretionary reviewer pays attention to the applicant, and attention is the scarce good.
The mechanism without that failure mode is the one where funding is decided retroactively, against measurable outcomes, by a large pool of independent voters with skin in the same game — and where the criterion is what shipped, not what was promised. I have not found a single instance of the null-report epidemic inside a program built that way. Not because those programs are staffed better. Because a report with nine null dimensions has no outcome to point at. It cannot survive a retroactive evaluation, because there is nothing in the past to evaluate.
The placeholder industry is, structurally, a forward-looking promissory industry. It can only exist where funding precedes delivery and nobody revisits delivery afterward. Change that ordering and most of the 371 disappear.
One more thing about the corpus, and then I will get to what I am watching.
I checked the temporal distribution. Null documents were not spread evenly across the 13 months. They clustered. The four highest-volume weeks in the entire corpus for placeholder reports were all within 72 hours of a major market event — a liquidation cascade, a large depeg scare, a regulatory headline.
That tracks with how I have seen newsrooms handle stress, including mine. When the tape moves, the instinct is to publish the shape of the analysis before you have the analysis. Nine headers, filled in later. Except "later" never arrives, because the next event is already here, and the null document is already published, and it is now a citable artifact.
I did exactly this once, in the other direction. May 2022, when TerraUSD came apart, I had eleven hours before my first publishable thread, and I filled them tracing the liquid staking derivative mechanics and the flash-loan sequence with two independent researchers — because the alternative was writing a shape I could not source. It was the right call and it cost me four hours against competitors who published first. Two of those competitors ran corrections within 36 hours.
Speed is the job. Speed with a verification floor is the job done properly. Speed without one is a null document with better prose.
So: what to watch. Three signals, all measurable, none of them requiring you to trust me.
Start with verification budget ratio — the share of a research allocation that goes to checking work rather than producing it. Below 5% and you are funding paper, not knowledge. Almost every program I checked sits below 1%.
Then citable null rate — the share of a governance forum's cited research that contains an empty dimension. This is one scraper away from being public data. I will be watching it monthly.
And reviewer-set overlap — the gap between who reviews and who signs. Under 40 points, and the "community" in community-verified is doing real work. Over 60, and you should count the fonts before you count the quotes.
None of the three requires a model to be smarter or an industry to become honest. They require someone to keep score. In a range, keeping score is the only edge there is.
The 22-page document is still open on my second monitor. Nine dimensions. Zero data. And somewhere, a treasury line item paid for it, a reviewer waved it through, and a vote used it.
I still do not know which vote. That is the part that keeps me up.