On the last Tuesday of February, a system I built refused to work, and I spent the better part of an afternoon trying to decide whether I was proud of it or furious with it.
The system is a research pipeline โ an unglamorous assembly of prompts, schema validators and retrieval calls that I put together for a small desk of analysts covering DeFi protocols. Its job is to take a document (a whitepaper, a governance post, a project's own breathless blog entry) and reduce it to a structured set of facts before any opinion is permitted to form. Nine dimensions: technical surface, token economy, market position, ecosystem dependency, regulatory posture, team and governance, risk matrix, narrative cycle, supply-chain transmission. I wrote that schema myself over three weeks, with the particular smugness of someone who has read a great deal of bad analysis and believes she has diagnosed the disease.
That afternoon I fed it a document and it returned, instead of an analysis, a page of refusals. Title: not provided. Source type: not provided. Core claim: empty. Information points: completely empty โ flagged fatal. Nine dimensions, each marked N/A, each carrying a short note explaining that the dimension could not be populated without facts the system did not possess. And at the bottom, in the flat register of a witness declining to testify: I will not fabricate an analysis that looks complete.
My first reaction was the ordinary irritation of a reader who wanted an answer and received a lecture. Within the hour, that irritation curdled into something less comfortable, because I recognised the shape of the document. It was mine. I had written that schema precisely to stop analysts from doing what the pipeline had just declined to do โ and I had spent four years watching humans, myself included, do it anyway.
What follows is not a defence of my software. It is an argument that the refusal it produced is the single most valuable artefact the crypto research industry could manufacture right now, and that almost nobody in this market is willing to pay for it.
There is a ritual that has grown up around crypto due diligence since the 2022 collapse, and like most rituals it began as a genuine prophylactic against death. In the years when Terra's mechanism was described in forty-seven pages of equations that almost nobody read to the end, when Celsius published yield tables with no visible asset side, when a lending desk's risk function consisted of a founder's conviction โ in those years the industry learned, expensively, that vibes are not collateral. The response was the checklist. Frameworks proliferated: token unlock schedules, team vesting cliffs, audit attestations, treasury runway, governance concentration. Consultants sold them. Funds demanded them. Grant programmes began attaching them to applications.
I am not sneering at checklists. I use them. A checklist is a memory aid for an analyst who already knows what the numbers mean; it becomes a forgery device the moment it is handed to someone who does not. That distinction is the whole essay, and it is almost never made.
Because here is what the bear market did to the ritual. When prices fall, the demand for certainty rises while the budget for producing it collapses. Funds cut research headcount first, because research is the only cost centre that never shows up in a mark-to-market. What remains is a smaller number of people producing a larger number of documents under greater time pressure, and a reading public โ depositors, LPs, small holders deciding whether to withdraw โ who want to know one thing: is my money safe here? They do not want nine dimensions. They want a verdict. The template is what you get when you promise a verdict you have not earned.
And then, in 2025 and 2026, the template met the machine. Anyone can now generate fourteen pages of plausible, well-typeset, correctly-titled diligence in ninety seconds. The marginal cost of the appearance of rigour has gone to zero while the cost of actual rigour โ an RPC endpoint, a Dune query, an afternoon spent reading a governance forum where twenty people argued about a parameter change โ has stayed exactly where it was. That asymmetry is the most important economic fact in crypto media, and it explains most of what you read.
I want to walk through what an empty report actually looks like from the inside, because I have now built one deliberately and watched it fail on purpose, and the experience taught me something about the epistemology of this industry that no amount of reading could have.
When my schema returned nine N/A values, it did not return nine identical things. It returned two very different classes of failure, and the distinction matters enormously.
The first class was absence of evidence โ dimensions that were unpopulated because the source document simply did not discuss them. Regulatory posture: not mentioned. Unlock schedule: not mentioned. These are gaps in the input, and they are cheap to close. You go and find the information; the chain, the forum, the filing, the auditor's PDF. An N/A here is a to-do item, not a conclusion.
The second class was absence of a source โ dimensions that could not be filled because there was no document to analyse at all. Here the N/A is not a gap in the input; it is a statement about the entire exercise. If nobody has told you what the asset is, there is no diligence to perform. There is only a request that you perform diligence anyway, so that the request appears to have been met.
Most crypto research fails in the second mode while presenting itself in the first. A report that says 'team: under investigation' looks like a report with a gap. A report that says 'team: anonymous, four of five wallets cluster to a single CEX deposit address funded from the same source in March 2023, governance quorum currently 61% held by three addresses' is a report that did the work. The first is a placeholder dressed as diligence. The second is a finding. They occupy the same row in the same table, and the visual similarity is the fraud.
I learned this โ properly, in the body rather than the head โ in 2018, auditing a donation contract for a DeFi prototype whose founders were earnest and whose code was not. The documentation described a withdrawal flow. The code implemented a slightly different withdrawal flow. The difference was one external call that arrived before the balance was decremented, and the gap between what the prose said and what the state machine did was worth roughly two hundred thousand dollars of other people's deposits. I have written about that contract before and will probably write about it again, because it taught me the only rule of verification I have never since improved on: read the state, not the prose. The prose is a claim about the code. The code is the code. When they disagree, the code wins, and the prose was never lying so much as describing a system that only existed in the author's imagination.
Everything I have said about templates follows from that rule. A research document is prose. The chain is the state. When a report says a protocol holds a certain amount of value, that is a claim about the state, and it is trivially checkable โ a contract address, a token balance, a decimal adjustment, done. When a report says the team is committed, that is a claim about a mental state, and it is not checkable at all, and it should therefore be marked as what it is: a bet, not a fact. The nine-dimension framework, properly used, is a machine for sorting claims into those two bins. Improperly used, it is a machine for making them look alike.
So let me be concrete about where the numbers live, because the vagueness of the previous paragraph is itself the disease I am describing.
The state is reachable. Any node will tell you the balance of any address; an indexer will tell you the history; a query layer over decoded logs will tell you the flows. This is, in the literal sense, public information โ the entire epistemic promise of the technology. What is not public information is the interpretation, and the gap between data and meaning is where the forgery happens.
Consider a number everyone quotes: total value locked. It is a sum over balances, and it is nearly meaningless without three qualifiers that are almost never printed alongside it. Which assets, at what prices, and โ crucially โ how much of it belongs to one address. A pool with nine figures of liquidity contributed by two market makers and a protocol-owned treasury is not the same object as a pool with nine figures contributed by four thousand depositors, even though the headline is identical. The difference is who leaves first in a drawdown, and it is the only question that matters in a bear market. I have watched a protocol lose two thirds of its deposits in a week while its TVL chart moved by less than a tenth, because the withdrawal came from a single concentrated LP whose position had been mispriced by the oracle. Nobody lied. The metric simply could not see what happened.
Or take a number that is quoted constantly and understood rarely: yield. There are only three sources of it in this industry. Real revenue paid to capital, emissions paid from a treasury, and someone else's principal. Every sustained yield above the risk-free rate is some blend of those three, and the blend is disclosed in the token contract, not in the marketing. You do not need a nine-dimension framework to determine which one you are buying. You need to look at where the payment originates โ an external fee stream, a mint function, or a deposit contract โ and then ask whether that origin is durable. That is a twenty-minute exercise and it would have saved an enormous amount of money in 2021.
The third number is the one that kills protocols in a bear market and the one least often surfaced: the unlock cliff. Vesting schedules are on chain and publicly derivable from the token contract, and yet I still receive analyses that discuss 'tokenomics' in the abstract โ allocation percentages, a pie chart, a sentence about long-term alignment โ with no date, no quantity, and no reference to the current float. A pie chart is not a schedule. A schedule is: this many tokens, on this date, to these addresses, of which this fraction has historically been sold within forty-eight hours of receipt. There are analytics platforms that will give you the first three of those for free. The fourth you infer from the addresses' history, and it is the one that determines whether the thing you are holding gets repriced next Thursday.
Now the important part. None of this is available to a language model reading a blog post. All of it is available to a person with a browser tab and an hour. The scarcity in crypto research is not intelligence and it never was; it is the willingness to spend an hour on a claim that will make you unpopular.
Which brings me to the Lightning Network, my favourite example of a topic on which the entire public conversation has decoupled from the state.
Lightning has been declared dead and declared inevitable, in the same week, every year, for seven years. The proclamations are made by people who have never run a node. If you run one โ and I have, badly, on a Raspberry Pi that is now a doorstop โ you discover that the interesting quantities are channel liquidity distribution, routing failure rates at various payment sizes, and the operational burden of maintaining inbound capacity. These are measurable. They are also unglamorous, and so the discourse discusses them approximately never. My own view, formed from that experience, is that Lightning will remain a useful settlement rail for a specific class of payments between well-capitalised nodes and will not become the retail payment layer it is repeatedly announced to be โ and note that this is a falsifiable claim about routing behaviour at scale, which is what makes it worth stating, unlike the ambient optimism and ambient doom that surround it.
The pattern generalises. Uniswap V4's hooks turn the DEX into programmable Lego, and anyone who has audited a hook understands that composability at that layer multiplies the audit surface by the number of hook combinations rather than by the number of hooks โ which is why I expect the overwhelming majority of hook deployments to remain trivial wrappers around fee logic while a small number of very sophisticated teams extract most of the value. That is not a criticism of the design; it is an observation about developer capacity, which is a measurable constraint. It can be argued with. It can be falsified by the deployment data. It is not a vibe.
And CBDCs. Here the industry's analytical habit fails completely, because the standard move is to compare feature sets โ a CBDC has programmable restrictions, so does a stablecoin, therefore they are on a spectrum. They are not on a spectrum. A CBDC is a ledger architected for observation; a bearer asset is a ledger architected for concealment. Those are opposite design objectives, and opposite objectives produce systems that cannot be averaged. You can build a payment rail that does some of each, and you will get a surveillance system with a privacy budget. You cannot build the midpoint between a panopticon and a vault, because the midpoint is a panopticon with better manners. I hold this view with high confidence and I state it as an argument rather than a measurement, and I try โ not always successfully โ to be honest about which of those two things I am doing.
That honesty is expensive, and the market prices it accordingly. Which is the economics of the thing, and it is worth spelling out, because pretending that the problem is individual laziness is a way of avoiding the actual mechanism.
Fabrication is cheap. Verification is expensive. A fabricated number costs nothing to produce, reads identically to a verified one, and is available now. A verified number requires an endpoint, a query, a sanity check and a timestamp. Under any plausible time preference, the fabricated number wins, and the only thing that ever punishes it is a reader who checks. Readers mostly do not check, because checking is exactly the expensive part, and so the equilibrium settles at confident noise with a thin crust of real work on top.
Bear markets sharpen this rather than softening it. In a bull market, being wrong is forgiven because everything goes up; the cost of a bad analysis is deferred indefinitely. In a drawdown, the audience is frightened, and frightened audiences do not want probabilities. They want a verdict, delivered in a confident voice. Certainty is the product with the highest margin in a falling market, and it is the one product nobody can actually manufacture. I have watched people I respect produce categorical statements about a protocol's solvency on the basis of a dashboard whose refresh interval they never checked.
I withdrew from public writing for six months during the 2022 crash, ostensibly because I was exhausted and actually because I had nothing honest to say. The token I had been associated with lost ninety-five percent of its value, my analysis of it had been wrong in ways I could not yet articulate, and every sentence I drafted about the market was either a prediction I did not believe or an apology I did not mean. What eventually got me writing again was teaching blockchain fundamentals to teenagers in Milan through a non-profit programme, where nobody asked me for a verdict and everybody asked me what a hash was. That experience rebuilt my relationship with the difference between knowing something and sounding like someone who knows something. They are not correlated in the way the industry assumes.
Now the part that I suspect will annoy the people who pay my grants.
The refusal to fabricate is not a virtue. It is a floor. My pipeline did not behave admirably; it behaved correctly, in the way that a calculator is correct when it declines to divide by zero. Celebrating it is a category error, and building an entire identity around analytical honesty โ as I have, professionally, for the last four years โ is one of the more insidious traps available to a person with my temperament, because it lets you feel principled while producing nothing.
Consider what an N/A genuinely accomplishes. It says: I do not know. Fine. But 'I do not know' is a starting position, not a deliverable. The document that landed in my inbox was honest and useless in equal measure. If I had been the reader paying for it, I would have been within my rights to ask why, given nine flagged dimensions, the analyst had not gone and filled three of them. Abstention that does not specify the path to knowledge is just pessimism with better formatting.
The honest version of that page would not have been nine N/A values. It would have been three things: what I could verify and how, what I could not verify and what it would cost to try, and what I would do with my own money under that uncertainty. Uncertainty with a price tag and a position attached. That is a document a human can act on. What I produced was a document a human can only admire โ and admiration, in this market, is a way of avoiding decisions.
There is a second, sharper version of the criticism, and it lands on my own tribe rather than on the template-industrialists. Decentralisation evangelists, of whom I am one, traffic in claims that cannot be falsified by data. 'This protocol empowers its community.' 'This design preserves sovereignty.' 'This unlocks human agency.' Every one of these might be true, and none of them can be refuted by a query. We built a vocabulary of values precisely because the numbers kept disappointing us, and then we mistook the vocabulary for evidence. The people producing confident nonsense about a token's price are at least constrained by a reality that will embarrass them on a schedule. The people producing confident nonsense about liberation are constrained by nothing at all. I include myself in that indictment without qualification.
Which is why I have spent the last year on a problem that looks, from the outside, like a category error and, from the inside, like the only problem left.
When I joined the SynthVoice campaign in 2026 to argue for verifiable human identity against synthetic media, the framing was: in a world of infinite generated content, cryptographic attestation of who produced a thing becomes the last redoubt of trust. The manifesto I wrote โ 'The Proof of Soul,' which is an embarrassing title that I stand behind โ made that argument in terms of dignity: a person's authorship should be provable by them and not by a platform that can revoke the claim. The project received half a million dollars in grants from foundations whose entire mandate is decentralisation, and I remain grateful, and I have come to think the framing was incomplete.
Because identity is the easy half. Proving that I am the author of a claim does nothing to establish whether the claim is true. A signed lie is a lie with better metadata. The real crisis is not provenance of people; it is provenance of facts โ the ability to trace a number back to the state, the query, and the moment it was read. And this is a problem that, uniquely, this industry is equipped to solve, because the underlying data is already public and immutable, and the missing piece is an attestation layer that most people have no commercial incentive to build.
I am not proposing a protocol. I have watched too many people propose protocols to solve epistemic problems, and they always produce a token, and the token always becomes the point. What I am proposing is a change in what a research document contains. Every factual claim carries its source, its method, and its timestamp. Not in a footnote โ in the body, structurally, so that a reader can re-derive the claim without trusting the author. A signature over the query, not over the conclusion. The conclusion is rhetoric, and rhetoric is allowed to be wrong. The query is not.
If that sounds heavy, consider the alternative, which we are already living. An AI writes the analysis from a document, another AI summarises the analysis, a third AI recommends the asset to a human who reads none of the three. The loop closes without any point of contact with the state, and the whole thing is fluent, coherent, internally consistent, and floating. That is not a hypothetical about the future. That is the current output of most crypto research pipelines, mine included, on days when I am not paying attention.
The empty page is the only artefact in that loop that tells the truth about itself.
The next bull market will not be won by whoever has the most data. Data is abundant and about to become free. It will be won by whoever can prove where their numbers came from โ a demonstrated chain from claim to state to timestamp, auditable by a stranger, on a bad connection, without asking permission. That is the actual decentralisation problem of the decade, and it is far less glamorous and far more consequential than anything currently being tokenised.
We have spent a decade learning to verify transactions and almost none of it learning to verify assertions. The first gave us a public ledger of movement. The second would give us a public ledger of belief, and I think it is the only infrastructure that makes everything else worth having.
So here is the question I now ask of every research document, including my own, including the ones that pay my rent: strip away the title, strip away the source, empty every field, remove the author's reputation from your mind entirely โ and ask what remains. If the answer is a structure of nine dimensions and no facts, you have learned something valuable about the document. If the answer is a handful of numbers with timestamps and a stated position, you have learned something valuable about the world.
The discipline is not in knowing more. It is in being able to tell, on demand and without flinching, which of those two things you are holding.