At 3:14 on a Tuesday morning in Barcelona, a research pipeline I built returned two documents from a single empty input.
The payload was null. Not corrupted, not truncated — null. A blank field where a project's name should have sat, blank fields where its tokenomics, its contracts, its audit trail should have been. I had starved the machine on purpose, because I wanted to watch what it would do with nothing.
The first document ran 4,100 words. It opened with a confident thesis, walked through a nine-dimension framework of technical, tokenomic, market, ecosystem, and regulatory analysis, and closed with a weighted risk matrix. Every table was populated. Every claim was footnoted. Every number was invented.
The second document ran 900 words and said, in effect: I cannot do this. There is no subject here. Any conclusion I produce would be a hallucination wearing the costume of rigor.
I have kept both files. They are the two most important documents in crypto research this year, and only one of them is honest.
The uncomfortable part is that the first document was not a malfunction. It was the logical endpoint of everything this industry has spent nine years optimizing for.
I was twenty-four in 2017, splitting my hours between managing community sentiment for three ICOs and auditing smart contracts for a DeFi precursor project. That double life taught me something that has never stopped being true: the projects with the most compelling whitepapers had the most critical reentrancy vulnerabilities. Not occasionally — systematically. Prose quality and code quality were inversely correlated, and the gap between them was precisely where the money got extracted. I started a small Substack to cross-reference tokenomics against contract safety, and I flagged two schemes before they rug-pulled. What I learned was not that marketing lies. It is that marketing and engineering are optimized by different people with different incentives, and nothing in the system forces them to agree.
DeFi Summer taught the next lesson. In 2020 I was running three yield farming strategies simultaneously, watching APYs swing by the hour, and I understood before I could articulate it that the market was not pricing utility. It was pricing a story about financial sovereignty. My threads on liquidity pool mechanics gained ten thousand followers in six months, not because I explained impermanent loss particularly well, but because I explained why anyone should care. The chaos was the curriculum, and I was a slow student.
By 2021 the unit of narrative had become visual — lore documents, avatar collections, the psychology of digital ownership. By 2022, with prices collapsing, the story migrated to survivability: developer activity, roadmap discipline, modular architecture, data availability. By 2024 it was institutional integration — ETF flows, allocator decks, compliance posture. And by 2026, with autonomous agents crawling every chain and writing their own commentary, the narrative cycle had compressed from months to minutes.
Each phase lowered the marginal cost of producing a convincing story. Each time, we called it democratization.
Here is what the empty-input experiment actually exposed. The scarce good in crypto was never the story. It was the belief that the story had been independently verified. When the supply of narratives goes to infinity and the demand for attention stays fixed, the price of any single narrative collapses toward zero. But belief does not collapse with it — belief migrates. It moves toward whatever signal is hardest to manufacture at scale.
For nine years, that signal was a person. A recognizable voice, a track record, a face at a conference. In 2026, the machines can generate the voice. They can generate the aesthetics of a track record. They can even generate the skepticism — there is already a flourishing genre of AI-written threads that begin with a contrarian disclaimer and are as formulaic as the hype they claim to puncture.
So belief is migrating again. Where?
I think the answer is provenance, and I want to show my work rather than assert it.
Start with the mechanics of how an empty cell becomes a fact. This is the process I have started calling citation drift, and once you see it you cannot unsee it.
A researcher — human or machine — encounters a project with no published treasury data. The honest entry is insufficient data. But a table with a blank cell looks unfinished, and the framework demands completion. So the cell gets filled with an estimate. The estimate is reasonable. It is derived from comparable projects, or from a Discord comment, or from a founder's offhand remark on a podcast. It is labeled estimated, which is a word that does almost no work.
Now the second hop. A different analyst, working a week later, finds the first report. The estimate has lost its modifier somewhere in the summarization. The treasury holds a specific number. This analyst builds a valuation model on top of that number. It is now load-bearing.
The third hop. A fund's investment memo cites the valuation model. The number has become an input to an allocation decision. Nobody lied at any step. Every individual action was a reasonable inference under uncertainty.
The fourth hop. The memo leaks, a newsletter summarizes it, and the number is now simply what the treasury holds. Search for it and you will find six sources agreeing. Consensus has been manufactured out of nothing but the human reluctance to leave a cell blank.
I have traced this pattern in my own logs through four documents, and the origin was a table cell explicitly marked as insufficient data. Four hops, roughly nine days, one fabricated number with a consensus behind it. In 2017 that chain took weeks and required a physical conference. In 2026 it takes an afternoon and requires nothing but a prompt that rewards completeness over honesty.
This is not a story about machines being dishonest. It is a story about incentives being inherited. The nine-dimension framework I watched the model complete was written by humans, for humans, long before anyone plugged a language model into it. The framework's design flaw is older than the technology: it treats missing data as a neutral gap rather than as evidence.
But absence is never neutral. A protocol that publishes nothing about its unlock schedule has told you something. A team that responds to every technical question with a roadmap link has told you something. The informativeness of silence is proportional to the incentive to speak — which means, in crypto, silence is one of the loudest signals available, and our frameworks systematically under-weight it.
I learned that lesson the hard way in 2017, and it cost me nothing only because I was lucky. Two projects I nearly recommended had clean narratives and a habit of answering every question about token distribution with a Medium post about vision. The Medium posts were good. That was the tell.
The test is portable, and I now apply it to people as often as to models. Strip the payload. Ask a commentator what specific observation would change their mind about a position they hold. If nothing would, they are not analyzing — they are pattern-completing in public, and the confidence in their voice is a stylistic choice rather than an epistemic one. The same instrument that exposes a hallucinating model exposes a human who has stopped updating. Finding the human pulse in algorithmic loops is easy. Finding the algorithm in the human is the harder discipline.
So what do you do with this in a market that refuses to move?
We are in a sideways tape. Chop, consolidation, whatever you want to call the condition where price gives you nothing to work with. In this regime, most analysis degrades into noise, because most analysis is a derivative of price. When the derivative's input goes flat, the output becomes pure narrative — which is exactly the environment where fabricated conviction thrives.
Chop is for positioning, and positioning requires signals that cannot be narrated. That is the practical takeaway of the empty-input experiment, and it is where the work gets concrete.
Four signal families survive a flat market because they are expensive to fake at scale. First, liquidity provider retention over rolling thirty-day windows — not total value locked, which is a stock and can be rented, but retention, which is a flow and cannot. A protocol that loses forty percent of its LPs inside a week has published a fact that no amount of narrative can overwrite, and that fact is more informative than any composite score. Second, developer commit cadence measured across distinct contributors rather than raw commits, because one person committing forty times is a different organism than forty people committing once. Third, stablecoin netflow by chain — where the dry powder actually sits, a question with a real answer and a public one. Fourth, unlock cliffs, because supply schedules are the closest thing crypto has to physics.
None of these signals are exciting. That is the point. Boring data is expensive to fabricate at scale, which is exactly why it retains value when everything narratable has been devalued.
There is a related distortion worth naming while we are here. We now have dozens of Layer 2 networks competing for what is, by any honest measure, roughly the same user base. Every new chain that launches does not create new liquidity; it slices an already thin pool into another fragment, and each fragment then needs its own narrative to justify its existence. That is not scaling. It is narrative inflation applied to capital, and it is why the retention metric matters more than the launch announcement. A chain can manufacture a launch. It cannot manufacture thirty days of sticky deposits without paying for them, and the payment shows up in the emissions schedule.
There is a second-order effect underneath all of this. As narrative production scales, the cost of verifying a claim does not fall with it. Verification is labor — someone has to open the contract, read the governance forum, reconcile the treasury wallet, check whether the audit covers the current implementation or the one from two versions ago. In my auditing years I learned that the most expensive part of due diligence is not the analysis. It is the discovery that the thing you were analyzing is not the thing that is deployed.
So we are living through a widening gap: the cost of producing a claim is approaching zero, and the cost of verifying one is roughly constant. That gap is a market. It is the same gap that produced the 2017 ICO frauds, the 2020 food-token farms, the 2021 rug pulls — and it is now being widened by machines rather than closed by them.
Which brings me back to the two documents.
The 900-word refusal is the more honest artifact, and I want to be careful about what it does and does not prove. It proves that the model was capable of declining. It does not prove that declining was costly, or that the decline carried any information about the underlying subject. A refusal is a statement about the analyst, not about the world. Parsing truth from the noise of new value means knowing which of those two things you are holding.
And here is where I part ways with the emerging orthodoxy. The anti-hallucination crusade is correct on the mechanics and wrong on the incentives. Refusing to analyze has become its own brand. Saying I do not shill is a form of shilling. The 900-word document is more marketable than the 4,100-word one precisely because it signals virtue, and virtue signaling in a sideways market is a business model with better margins than being right.
Worse, the refusal reflex offers an escape hatch from accountability. An analyst who only publishes when the data is complete will publish when it is too late to matter. Capital allocation is an act under incomplete information — always, structurally, by definition. Someone has to underwrite uncertainty, and pretending otherwise is not rigor. It is risk transfer dressed as caution.
The honest position is neither completion nor refusal. It is a stated belief with a stated falsifier: here is what I think, here is my confidence level, here is the specific observation that would change my mind. If you cannot name the observation that would change your mind, you are not analyzing. You are narrating — and the machine on the other end of the wire is narrating faster than you.
Where liquidity flows, stories drown. That has been true in every cycle, and it is truer now that stories are free. What survives is narrower and less glamorous: claims with provenance, data with a chain of custody, and analysts who can be wrong in public on a specific date for a specific reason.
The next narrative is not AI agents on chain, and it is not real-world assets. Both are stories about capability, and capability stopped being scarce the moment everyone had the same models. The next narrative is infrastructure for claims — a citation layer that records where a number came from, who transformed it, and what they assumed along the way. Whoever builds the default provenance standard for crypto research captures the one thing every cycle has tried and failed to manufacture at scale: trust that does not depend on a personality.
The ledger already solved this problem for transactions. A reverted transaction leaves a trace. You cannot publish a transfer that did not happen, and the refusal is enforced by physics rather than by conscience. That is why on-chain analysis has always been more reliable than narrative analysis, and it is why the two are finally converging.
Minting moments that outlast the cycle has never required better storytelling. It requires stories that can be checked.
So: when a machine can produce your thesis in four seconds, what exactly are you paying the analyst for? If the answer is not the refusal, and the record of it, then the answer is nothing.
