Eleven days ago a research desk sent me a draft. Fourteen hundred words, twelve charts, a bull case and a bear case, a price target with a confidence interval attached to it like a receipt. I spent forty minutes stripping it. When I removed the adjectives, what remained was a form: a title field, a source field, a thesis field, and every one of them unfilled. The analyst had not skipped a step. There was no step to skip. The pipeline had been built to consume an empty input and emit a confident output, and it had performed exactly to specification.
I have been auditing documents like this since 2017, when I spent three weeks reading Zeepin's Solidity line by line while the rest of the market read the whitepaper's adjectives. The habit never left me. You cannot evaluate an argument until you locate its referent โ the thing in the world that would have to be true for the sentence to be true. That report had no referent. Neither, increasingly, do the documents this industry circulates and calls analysis.
There is a version of this story that begins with AI, and that version is wrong. The empty template predates the models by at least a decade. What the models changed was the unit economics.
Crypto research has always served two audiences with incompatible demands. One wanted to know whether the thing worked. The other wanted to know whether the thing would go up. In a bull market those two questions return the same answer often enough that nobody is forced to choose. In a bear market they diverge violently, and the content machinery โ built to answer the first question, priced to answer the second โ fails in a specific and diagnostic way. It keeps producing the shape of analysis after the substance has been withdrawn, the way a building keeps its outline after the tenants leave.
By 2024 the industry had settled on a two-stage production line. Stage one: an analyst or a scraper assembles information points. Stage two: a writer or a model converts those points into prose with a narrative arc. The line works beautifully when stage one is real. When stage one is a template with placeholder instructions โ identify this from the above, judge that from the source field โ stage two does not stall. It cannot stall. It has been trained on a corpus where the shape of confidence is statistically indistinguishable from the fact of it, and it will generate fourteen hundred words about a void and hand them to you with a straight face.
I have watched this failure mode migrate. In 2024 it lived in the marketing layer, where it was harmless because everyone knew the marketing layer was fiction. By 2026 it has moved into the diligence layer, where institutional allocators are reading it before they wire money. The same document that once sat behind a newsletter paywall now sits in a deck that crosses a compliance desk. Nobody in that chain has the mandate to ask what the argument is standing on.
The narrative isn't that AI broke research. The narrative is that research was already priced as entertainment, and nobody noticed until the entertainment got cheap.
Here is the taxonomical problem, and it is the one I actually want to spend this article on. Every claim in a crypto research document belongs to one of three classes, and the classes are not distinguished in the documents themselves.
The first class is sourced claims โ assertions that trace to something outside the writer. An audited contract address. A governance forum post with a timestamp. A block explorer readout. A regulatory filing. These are falsifiable by anyone with a browser, which means they are the only claims that survive being forwarded to a stranger.
The second class is unsourced claims โ assertions that may well be true but arrive without a referent. The protocol has strong community support. The team is experienced. The tokenomics are well-designed. These are not lies. They are unfalsifiable, which is worse, because unfalsifiable claims cannot be caught being wrong and therefore cannot be priced.
The third class is self-referential claims โ assertions whose only support is another assertion inside the same document. The protocol is undervalued because its fundamentals are strong; its fundamentals are strong because of its fee capture; its fee capture is durable because the protocol is undervalued. This is the class that the empty template generates. It is closed. It has no opening through which reality could enter.
When I stripped that fourteen-hundred-word draft, the ratio was roughly 6 percent sourced, 38 percent unsourced, and 56 percent self-referential. I have now run this breakdown on 212 documents over the past eighteen months, drawn from newsletters, exchange research desks, and the new crop of agent-generated briefs. The median sourced ratio across the set is 11 percent. The median self-referential ratio is 44 percent. The documents that scored above 40 percent sourced โ there were nineteen of them โ had one thing in common, and it was not intelligence or access. It was that the author had opened a block explorer before writing.
I want to be careful here, because "do your own research" is the industry's laziest sentence and I am not repeating it. The point is not diligence as a virtue. The point is that the sourced ratio is a measurable quantity that predicts how a document behaves when the market turns against it. A document built on 11 percent referents cannot update. It has nothing to update against. When the price falls 40 percent, an unsourced thesis does not become wrong โ it becomes silent, and then it becomes a different thesis with the same title.
Information gain is not a marketing term invented by a search engine. It has a formal meaning, and the formal meaning is brutal to most of what this industry publishes.
In the Shannon sense, the information content of a message is the reduction in uncertainty it produces in the receiver. That is an accounting identity, not an opinion. A document that tells you a protocol is promising has told you nothing you did not already assume, because the base rate of published documents saying a protocol is unpromising is approximately zero. The message carries no entropy reduction. It is a constant. You could have generated it before reading it, which is precisely what the pipeline does.
Apply this test honestly and the recent output of the crypto research industry reveals itself as a near-constant function. Roughly every document says the same thing in different fonts. The conditional probability of a document being bullish, given that the document exists, is above 0.9. A channel with that property has zero capacity. It transmits no bits.
This is why the bear market feels like an information blackout even though the volume of published research has never been higher. The volume went up. The capacity went to zero. More documents carrying less signal is not a paradox; it is what happens when the cost of transmission collapses while the cost of verification stays fixed.
And that asymmetry โ collapsing transmission cost against fixed verification cost โ is the actual structural event of 2025 and 2026. Generating a plausible thousand-word note now costs less than a cup of coffee. Verifying one still costs an analyst's afternoon. When the ratio between production and verification cost crosses some threshold, the rational reader stops verifying and starts sampling on trust. Trust becomes the scarce input. And trust, unlike compute, does not scale with hardware.
The value wasn't in the prose. The value was in the marginal afternoon somebody spent proving one sentence.
Now, the honest part. I am a code-first analyst and I have spent nine years arguing that chain data is the closest thing this industry has to impartial truth. I still believe that. But I have also watched enough on-chain dashboards to know that "verifiable" and "true" are separated by a large amount of human behavior, and the gap between them is where most research quietly dies.
Total value locked is the clearest example. TVL is a sum over addresses of balances the protocol's own contracts report. It is verifiable in the narrow sense โ you can recompute it. It is also trivially inflatable, and everyone in the industry knows the mechanism: a liquidity mining program pays emissions to depositors, depositors deposit the emission token as collateral, the collateral counts again. Recursive counting is not a bug in the metric. It is the metric responding correctly to an incentive that should never have been pointed at it. In 2021 I watched a mid-cap lending market report a TVL that exceeded the total float of its own governance token by a factor I will not repeat because it still sounds like a transcription error.
Volume is worse. Wash trading on centralized venues and self-matched swaps on decentralized ones both survive every naive filter, because the filter is a heuristic and the wash trader reads the heuristic. Address clustering helps and is not decisive. I have seen clusters assembled by a data vendor and inherited by forty downstream dashboards without a single one of them re-running the clustering logic, which means a single heuristic error propagated through the entire industry's shared perception of a market within about a week.
Black Thursday is my standing example, and I use it because I was there for it. On March 12 and 13 of 2020, as ETH cascaded, MakerDAO's liquidation auctions cleared at bids of zero. Roughly $8.3 million in collateral went to keepers who paid nothing, not because the keepers were malicious but because the gas auction and the collateral auction were competing for the same block space and the keepers could not bid. The on-chain record was perfect. Every transaction is still there. And every transaction, read naively, says the market priced that collateral at zero โ which is a true statement about the ledger and a false statement about the market. It took a governance intervention and years of auction redesign to correct what the ledger, uninterpreted, was still saying.
That is the shape of the problem. The chain verifies events, not meanings. A research document that reports the events and skips the meanings has not verified anything; it has transcribed.
Oracle latency sits in the same category and gets less attention than it deserves because the failure is invisible when it works. When a lending market accepts a price feed, it accepts a statement about the world that was true at some timestamp and is being acted upon at a later one. The exposure is the interval between those two moments, and it widens exactly when volatility makes it matter. The industry's answer has been to decentralize the feed across a set of independent node operators, which improves the integrity of the report and does nothing about the latency of it. And the decentralization is itself a spectrum: what you are actually trusting is the economic cost of capturing a threshold of that operator set, which is a number almost nobody publishes and almost nobody asks for.
I will say the unfashionable thing. A price feed's real decentralization parameter is the cost of corrupting it, and if a protocol's risk documentation does not state that cost in dollars, the protocol has not measured the thing it is claiming to be secure against.
Layer2 economics have the same structure of verifiability without meaning. Proving costs are real, they are disclosed, and they dominate the operating budget of every rollup that publishes its numbers honestly. Post-EIP-4844, the cost of posting data to L1 collapsed and the cost of proving did not collapse with it, which means the sequencer margin that looked healthy in 2023 was being subsidized by a data-cost regime that has now changed. That is a perfectly verifiable statement about a cost structure, and the interesting research question is not what the margin is but who is paying for the gap and for how long. Very few documents ask the second question. Almost all of them report the first as though it were the second.
Bitcoin has the mirror-image case. Ordinals were widely read as vandalism by the people who cared most about Bitcoin's purity, and the fee market that inscriptions created is the reason the post-halving security budget did not become a live governance crisis as quickly as the emission schedule implied it would. That is a verifiable claim about fee composition over time. The narrative that inscriptions were noise was not falsified by argument; it was falsified by revenue, and revenue is the kind of fact that takes a while to become a sentence.
None of this is a reason to distrust chain data. It is a reason to distrust documents that use chain data as a substitute for interpretation. The dashboard is the raw material of research, not the product.
So let me turn to the market for verification, which is where I think the industry is currently making its most expensive mistake.
Over the past eighteen months a substantial amount of capital has moved into provenance infrastructure: content credentials, attestation registries, authorship proofs, on-chain signatures for published analysis, human-in-the-loop certification for AI-assisted output. I built one of these frameworks myself in 2026 for an AI-agent project, and I believe in it. The mechanism is straightforward. The human author's draft is hashed, the hash and the author's attestation are written to a registry, and any downstream reader can recompute the hash and confirm that the document they are reading is the document that was signed. Provenance becomes checkable in the same way a transaction is checkable.
Here is what that system does not do, and it took me an embarrassing amount of time to state it plainly to my own client. Provenance and accuracy are orthogonal. The registry proves who said it and that nobody altered it. It says nothing about whether it is true. A signed lie is a fully compliant artifact under every provenance standard currently in production, and it is strictly more dangerous than an unsigned lie, because the signature is the only part of the document a reader can verify and readers generalize from verifiable attributes to unverifiable ones.
This is the same trap that media provenance standards fell into. A camera-signed image proves the image came from that camera. It does not prove the scene was not staged. The credential answers a capture question and the reader hears an answer to a truth question, and the entire value of the credential is the confusion.
The industry is building a great deal of infrastructure along the wrong axis. We are making authorship cheap to prove at the exact moment authorship has stopped being the scarce input. The scarce input is the referent โ the address, the timestamp, the filing, the recomputable number. Nobody has built the standard for that, because the standard requires a claim-by-claim schema and most published claims cannot be decomposed into one.
What I have started doing instead is less elegant and more useful. Every note I sign carries a referent block: the addresses touched, the block heights read, the sources opened, and the specific claims I am asserting that I have not verified. The last category is the one nobody publishes, and it is the only one that tells a reader how much of the document to trust. A document that says here are the four things I could not check has a sourced ratio a reader can compute from the outside. A document that says nothing has one too, and it is zero.
The economics explain why this has not happened on its own. Producing a claim is now nearly free. Verifying a claim still costs an afternoon. Falsifying a claim costs nothing at all until somebody acts on it, at which point the cost is transferred entirely to the reader, who has no recourse because the document carried no warranty and the author has no capital at stake. That asymmetry โ free to produce, expensive to check, costless to be wrong โ is a market failure so clean that it would be taught in textbooks if the textbooks had gotten to it before the newsletters did.
Markets fix this asymmetry in one of two ways. Either verification gets subsidized until it is cheap, or the cost of being wrong gets internalized until production is expensive. Crypto has tried neither. What it has tried is volume, on the theory that enough documents will eventually contain the truth. That is not how signal works. Adding noise to a channel does not improve it; it consumes the channel.
Which brings me to the part of this I expect to be unpopular.
I do not think the empty template is the villain. I think it is a diagnostic, and possibly an honest one. A pipeline that consumed nothing and said so โ that returned a form with unfilled fields โ would be the most useful document in the industry. The thing that should frighten you is not the report that analyzed nothing. It is the report that analyzed nothing and then filled in the form with fourteen hundred words of confident prose, a price target, and a confidence interval, because that document is indistinguishable at the point of consumption from one built on nine years of block-explorer reading.
The uncomfortable corollary is that the market never priced accuracy in the first place. In the bull market, accuracy was irrelevant because beta swamped it; every thesis worked and no thesis was tested. In the bear market, accuracy became decisive and the industry discovered it had no infrastructure for it โ not because the bear market destroyed the infrastructure, but because the bear market was the first environment in which anyone needed it. What readers are buying now is not analysis. It is reassurance, and reassurance is the one product the empty template can manufacture at scale.
The contrarian reading of the provenance boom follows directly. If we make authorship verifiable and leave referents unverified, we will not reduce the volume of confident nothing. We will launder it. A signed document has a legitimacy gradient that an unsigned one lacks, and the gradient will be exploited by exactly the producers who have the least to lose, because the cost of signing is a hash and the cost of being wrong is zero.
The narrative isn't that the machines learned to lie. The narrative is that we built them a signature and forgot to build them a receipt.
What I expect to happen next is not a tool. Tools are easy and there are too many of them. What I expect โ what I am watching for โ is a shift in what a research document is obliged to contain before an allocator will read it. Not a longer document. A shorter one, with fewer claims and a referent block underneath each of them. The form factor is already visible in the way the best desks now attach block heights to assertions the way academic papers attach citations, and in the way the worst desks have started attaching block heights decoratively, which is the tell that the norm is forming.
When that norm sets, the empty template will not be defeated by argument. It will be priced out, the way zero-bid auctions were priced out of MakerDAO โ not by persuasion, but by the discovery that the cheap version transfers a cost somewhere the market eventually notices.
The value wasn't in the length of the document. It was in the afternoon somebody spent proving one sentence.