I want to start with a file — not a whitepaper, not a token announcement, not a governance proposal with a suspiciously round quorum. A research file. It landed on my desk in Jakarta eleven days ago, generated by an automated pipeline that a mid-sized trading desk had paid a five-figure annual license to run. Forty-one pages. Nine analytical dimensions: technical architecture, token economics, market structure, ecosystem niche, regulatory posture, team and governance, risk surface, narrative durability, supply-chain transmission. Every cell in that document read the same four characters — N/A.
The head of research forwarded it with a single line: "Is this thing broken?"
I told him it was the only piece of crypto research he had received that month that had not lied to him. That sounds like a joke. It is not. In the loudest bull market I have watched in twenty-nine years of orbiting this industry — a market where a freshly funded Layer 2 with $100 million in the bank, no mainnet, and eleven GitHub commits gets three "deep dives" published about it before lunch — the scarcest commodity is not alpha. Alpha is everywhere, or at least the claim of it is. The scarce commodity is the sentence: I don't know, and I am not going to pretend otherwise.
This is a piece about that sentence. About why it has become so expensive to say, why the machinery of modern crypto research is structurally biased against saying it, and what an empty file taught me about the difference between analysis and performance.
The supply chain nobody audits
Start with the obvious question: where does crypto research actually come from in 2026?
A decade ago the answer was simple and slightly embarrassing. A handful of analysts with too much time and a Twitter account read the code, read the forum, read the token distribution table, and wrote down what they found. The output was slow, uneven, and frequently wrong — but it was traceable. You could argue with the analyst. You could ask where a number came from, and the analyst would either produce the source or admit they had guessed.
That supply chain has been rebuilt from the bottom up. Today the pipeline looks like this: a protocol announces something — a mainnet date, a points program, a token generation event, an integration with a name you recognize. Within hours, a content layer spins up: automated summarizers, LLM agents with browsing tools, newsletter farms staffed by three people and forty prompt templates, and a long tail of "research" accounts whose business model is volume. Within a day, that content is aggregated into dashboards. Within a week, it is cited as background in a genuine institutional memo.
Nothing in that chain is audited. Nothing in it is required to be true. And the incentive at every step points in the same direction: produce more, faster, and with more confidence than the person next to you.
I came into this industry through the code side, and the code side taught me the opposite instinct. In 2017, still an academic with a mathematics background and a lot of free evenings, I was auditing early Solidity contracts for a DAO precursor project called EtherHouse. I found four re-entrancy vulnerabilities in the pre-sale logic — the classic external-call-before-state-update pattern, three of them exploitable in sequence. The team patched before the sale went live. Roughly $200,000 in pre-sale funds stayed where it belonged. That experience did something permanent to my brain: it taught me that the most dangerous finding in an audit is the one you invent. A false positive does not just waste time. It burns the credibility you need for the true positives.
That instinct — the auditor's instinct — is almost entirely absent from the research layer today. And in a bull market, its absence is not a small inefficiency. It is the primary mechanism by which capital gets misallocated.
How a framework becomes a hallucination engine
Let me get mechanical for a moment, because the failure mode here is not mysterious. It is a property of the tools.
A large language model generates text by predicting, token by token, what plausibly comes next. Give it a blank page and a vague instruction and it will produce something reasonable and generic. But give it a structured template — a nine-dimension table, a due-diligence checklist, a framework with labeled boxes — and you have done something much more specific. You have created an extraordinarily strong prior for completion.
The header says "Technical Innovation Score." The next token is almost certainly a number. The header says "Competitive Differentiation." The next tokens are almost certainly a comparative phrase. A table is a machine for generating confidence. Its shape communicates that every cell has an answer; the model, being a pattern-completion engine rather than a truth engine, obliges.
This is why the empty file mattered. Whoever — or whatever — produced that forty-one page document had been given the same framework as everyone else. But somewhere in the pipeline there was a gate that checked whether the underlying inputs existed. There were no inputs. No code repository. No token contract. No verifiable team history. No on-chain footprint. And so the gate closed, and the document came out full of N/A.
That gate is the entire product. Not the framework — anyone can write a framework. Not the prose — the prose is the cheapest part. The gate is the thing you are paying for, and almost nobody is paying for it.
I have spent the last two years running a hybrid education platform out of Jakarta — online courses paired with physical workshops, two hundred developers and a thousand business leaders trained through it since 2024. The single hardest thing I teach is not Solidity, not consensus, not regulatory mapping. It is the discipline of distinguishing "I could not verify this" from "this is false." Those are completely different statements, and in a market that rewards decisiveness, conflating them is the default behavior.
The anatomy of a number nobody can source
Let me make this concrete, because abstraction is how the problem hides.
Take the Lightning Network. For seven years now, a specific category of statistic has circulated through research decks: routing failure rates. You will see figures — percentages, decimal points, confidence intervals implied by the precision — presented as evidence about whether Lightning can scale. I have asked, politely and then less politely, for the methodology behind several of these numbers. The answers are always some version of: it was measured once, on a specific node set, under specific channel conditions, by someone who did not publish the raw data, and it has been re-quoted ever since.
What the honest version of that analysis looks like is different. It looks like: channel management is an operational burden that most users will never accept, inbound liquidity is a market that requires capital and attention, and the resulting system serves a real but narrow set of use cases. That is a defensible position built on observable mechanics. The routing-failure percentage is a costume that position wears when it wants to sound like data.
The same disease shows up in the data availability conversation, and here the gap between the number and the reality is almost comic. The current cycle is full of research on DA demand curves, DA fee markets, DA competitive dynamics — entire frameworks built on the premise that rollups are generating enormous volumes of data that need somewhere to live. Then you open a blob explorer and look at actual utilization.
The overwhelming majority of rollups are using a rounding error of the blob space available to them. Not 50%. Not 20%. A fraction of a percent, with occasional spikes during airdrop claim events. The DA layer is not a bottleneck. For almost every rollup in production, it is an empty parking garage with a very sophisticated pricing model.
You cannot write that sentence if your template demands a "DA Strategy Assessment" cell. But you can write it if you have looked at the explorer.
And then there is the hooks question — Uniswap V4, the programmable pool architecture, the thing everyone agrees is the future of the DEX. The technical claim is real: hooks turn a liquidity pool into a composable primitive, and that is genuinely a different design space. The question a template will not ask is the adoption question. How many hooks are actually deployed, in production, with meaningful liquidity, written by teams who understand the gas and security implications of arbitrary code executing inside a swap path? The answer is a small number, and it is small for a reason that has nothing to do with marketing: the complexity ceiling arrived before the developer base did. A design space that requires you to reason about re-entrancy, MEV, and hook ordering simultaneously is not a design space that ninety percent of teams can ship in safely.
None of these three examples is a scandal. They are all just cases where the honest answer is unflattering, and the research layer has a structural allergy to unflattering answers.
What ground truth actually looks like
So what is the alternative? It is not cynicism. Cynicism is just hallucination with a negative sign.
The alternative is building research on top of things that can be checked by someone who does not trust you. In practice, that means a small number of sources that share one property: they are generated by the protocol itself, not by a human describing the protocol.
Contract bytecode and verified source. Transaction history. Token holder distribution, computed from the chain rather than from a dashboard's summary of the chain. Governance proposal records, including who voted and with how much. Validator or sequencer sets and their actual composition over time. Bridge flows. Blob space utilization. Client diversity statistics. These are not opinions. They are readings.
Everything else — team quality, roadmap credibility, competitive positioning, narrative durability — is interpretation layered on top. And interpretation is legitimate. I do it constantly. But interpretation must be labeled as interpretation, and it must be falsifiable, and it must be anchored to readings that anyone can reproduce.
The distinction matters most in exactly the situations where it is hardest to maintain: when the price is moving. In 2022, after the Terra collapse took the market apart, I spent three months in my Jakarta apartment doing nothing but reading. Not trading. Reading. I produced a fifty-page dissection of algorithmic stablecoin design, and the core of it was a distinction that took me those three months to articulate cleanly: the difference between cryptographic trust and economic confidence.
A cryptographic guarantee holds under adversarial conditions by construction. An economic guarantee holds as long as the incentive structure holds, and the incentive structure holds as long as growth continues. Terra's design was presented as trustless. It was not trustless; it was confidence-dependent, dressed in the grammar of mathematics. The reason that analysis resonated — it circulated widely among people who had lost money — is not that I was clever. It is that I refused to accept a number I could not derive.
That refusal is the whole method. And it is expensive, because it means writing reports that say less than your competitors' reports say.
The classroom experiment
Here is the finding that unsettled me most, and it came from teaching rather than from auditing.
In one of my BlockJakarta workshops, I run an exercise. I hand out a token I invented. I call it SelatCoin. I give it a plausible whitepaper — a modular settlement layer, a points program, a partnership with a fictional regional payment processor, a tokenomics chart with a 4-year vest and a suspiciously small community allocation. I give the cohort ninety minutes and ask for a due-diligence memo.
Roughly eighty percent of participants produce a confident verdict. Not a hedged verdict — a verdict. Several of them assign scores. A few recommend allocation sizes. The token does not exist. There is no contract, no team, no chain, no code. There is a PDF I wrote on a Saturday.
The participants are not stupid. Many of them are working engineers and analysts. What they are is trained — trained by an information environment in which a document with structure is treated as a document with substance, in which the absence of a red flag is read as the presence of a green one, and in which saying "I could not verify any of this" feels like a failure to do the job.
That last part is the crux. The reason people fill in the blank is not that they believe the answer. It is that leaving the blank empty feels like admitting they did not do the work.
I recognize the feeling. In 2020, during the DeFi Summer chaos, I forked three different automated market makers in a co-working space in South Jakarta and launched UniBarter, a localized AMM aimed at Indonesian traders. It attracted five hundred users in two weeks. And then it ate my life — maintenance, oracle edge cases, liquidity incentives, the endless operational grind — until I understood that I was building infrastructure I could not sustain and solving problems I had not chosen. I shut it down and pivoted to teaching.
What I learned from that failure is the same thing the classroom experiment teaches: the pressure to appear productive is stronger than the pressure to be correct, and it operates from the inside. It is not a market force imposed on analysts. It is a habit of mind, and it is the habit that automated research pipelines were built to satisfy.
The economics of a confident lie
Now the uncomfortable part: this is not a bug. It is a business model.
Hallucinated research is cheap to produce and expensive to disprove. A confident forty-page report costs a few dollars of inference and an hour of prompt engineering. Disproving it costs a week of primary-source verification, and the person doing the disproving gets a fraction of the attention the original received. The asymmetry is brutal and it compounds: the correction is written a week later, when the market has moved on and nobody is reading.
This is why "information gain" — the requirement that a piece of content contain at least one thing the reader did not already know — is such an awkward standard for the current environment. Most crypto research does contain information gain. It is simply new rather than true. A fabricated fee-revenue projection is, technically, information the reader did not have before. That is the trap. Volume and novelty are cheap. Verification is the only expensive input, and it is the only one the market does not pay for.
There is also a deeper irony worth sitting with. This industry's entire philosophical proposition is verifiability — don't trust, verify; the chain is the source of truth; consensus replaces the intermediary. And yet the layer that produces the industry's own self-understanding has quietly centralized onto a handful of model providers, a handful of data aggregators, and a handful of newsletters that everyone cites without reading. We rebuilt the trusted third party, and we put it in the research stack.
I am not making an argument against these tools. I use them daily. I am making an argument about where the gate goes. If the tool that generates your conclusions is the same tool that summarizes your evidence, you have not built an analysis pipeline. You have built an echo with better typography.
The contrarian read: the empty report is the product
Here is the angle I keep coming back to, and it is the one I think is genuinely mispriced.
Everyone treats a negative result as a failure of research. Nobody funds it. Nobody shares it. A report that concludes "this project cannot be evaluated with available information" reads as a non-answer, and in a market that pays for conclusions, non-answers do not get paid.
But flip it. In a bull market, the population of assets that cannot be evaluated is not noise — it is the single most valuable output the research layer could produce, because it is the exact complement of the population that can. If eighty percent of what is being marketed to you is unverifiable, then a system that reliably identifies the verifiable twenty percent has an edge that no amount of faster summarization can replicate. The negative result is not the absence of alpha. It is the shape of the alpha.
There is a second, harder version of this argument. We have spent a decade building an industry around the promise that decentralization removes the need for trust in intermediaries. What we have not done is extend that promise to knowledge. We still trust the analyst. We still trust the aggregator. We still trust the model. And the model, being a completion engine, has a structural incentive to be fluent rather than correct — which is precisely the failure mode that decentralized systems were designed to eliminate from finance.
So the real question is not whether AI can do crypto research. It can, and it will, and much of it will be good. The real question is whether we will build the verification layer to match — reproducible data pipelines, source-linked claims, explicit confidence labels, and a cultural willingness to publish a page full of N/A and call it a finished product. That is a much harder engineering problem than prompting, and it is the one nobody is racing to solve, because the reward for solving it is a document that looks, on the surface, like it does nothing.
From core dev trenches to community heartbeat — that was the arc of my own career, and I think it is the arc this research layer has to walk too. The trenches taught me that a finding must be reproducible. The community is now teaching me that reproducibility is not a technical standard. It is a social one, and it has to be taught, workshop by workshop, analyst by analyst, until saying "I don't know" stops sounding like an apology.
What I am watching next
Education is the new mining rig for the mind, and right now the rig is pointed at the wrong ore. We are hashing attention into confidence and calling the output knowledge.
When the market sleeps, the architects wake up. The next cycle's most valuable research infrastructure will not be the fastest summarizer or the prettiest dashboard. It will be the gate — the unglamorous, unshareable, unsponsored mechanism that decides whether a document gets written at all. Build the gate. Everything else is already commoditized.
And when the next forty-one page report lands on your desk, ask one question before you ask whether it is broken. Ask how many cells said N/A — and whether anyone had the nerve to leave them that way.
