At 2:14 a.m. Shanghai time, a report landed in my inbox with nine sections, eleven tables, and not a single fact in it.
I had spent the evening watching an automated research pipeline do what much of the industry now outsources to machines: crawl a news article, extract its claims, tag its domain, and hand the residue to a second model for what everyone politely calls deep analysis. The first stage returned nothing. Not an error. Not a timeout. An empty structure — title not provided, source not provided, information points: zero, domain tag: unclassified.
The second-stage model, to its credit, refused to invent anything. It marked thirty-odd fields insufficient. It declined to name a project, declined to sketch a token supply chart, declined to estimate price impact, and closed with a short list of what it would need before it could begin. No narrative. No thesis. No numbers pulled from the air to fill a table.
I have read a great deal of machine-generated crypto research over the past eighteen months. That empty report is the most honest document I have received all year.
Crypto research has become an industrial process, and like all industrial processes it optimizes for throughput. A mid-sized desk can publish forty market notes before breakfast. Newsletters get assembled from scraper output. "Protocol deep dives" get generated from a whitepaper PDF and a price page. Sentiment dashboards aggregate bot traffic and label it alpha. The economics are unforgiving: reader attention is finite, the supply of analysis is infinite, and the only metric that scales is volume.
I understand the temptation, because I have been on the other side of it. In 2022, as a final-year student watching FTX and Celsius come apart, I spent six months auditing the economic models of failed projects and published a series I called Anatomy of a Collapse. My peers were pivoting to traditional finance. I stayed. What I learned in those six months shaped everything I have written since: those failures were rarely cryptographic. They were failures of incentive design and, more often, failures of honesty. Concentrated power creates moral hazard, and moral hazard generates documents — attestations, proof-of-reserves, "independent reviews" — that say precisely what someone needed them to say.
The industry learned a version of that lesson about money. It has not yet learned it about information.
Three years on, with a master's in applied mathematics and a desk in Shanghai, I designed incentive models for a Layer 2 and learned the corollary: an elegant mechanism with no social adoption is a museum piece. Efficiency without legitimacy is hollow. I now believe the same thing about analysis. Precision without provenance is hollow. If you cannot say where a number came from, the number is decoration.
A research pipeline looks simple from the outside. Inside, it is a chain of assumptions, and every link can fail quietly. The crawler fetches a page. The parser strips navigation, ads, and boilerplate. An extractor pulls discrete information points — the smallest claims that can stand alone and be checked. A tagger assigns domain and category. Only then does the analyst model see anything at all.
The failure mode that matters is this: at the schema level, an empty array is a valid array. The pipeline reports success, because nothing threw an exception. Nobody looks, because nothing on the dashboard is red. The most dangerous output a data pipeline can produce is not a wrong answer but a well-formed empty one — the schema validates it and the human never opens it. I have watched the same pattern in DeFi. A liquidation bot running on a stale oracle does not error out; it keeps quoting. Silence looks exactly like calm.
The empty report was not telling me the market was quiet. It was telling me the plumbing had been cut. A reader who cannot tell those two states apart will eventually pay to learn the difference.
Consider what a real information point looks like, because granularity is the whole game. A sentence claiming a protocol "is gaining traction" is not an information point. It is a mood. A sentence stating that a protocol's daily active addresses moved from one figure to another between two named dates, sourced to a specific dashboard, is an information point — because somebody else can check it. Everything downstream of extraction inherits that standard. Feed a model moods and it will return moods with more confidence.
There is also the dating problem. The pipeline's own schema carried a field for time sensitivity, and it came back unevaluated. That omission matters more than it sounds. A governance proposal's status flips the moment a vote closes. An exploit disclosure rewrites a protocol's entire risk profile within a single block. Undated analysis is not analysis; it is a horoscope with a chart.
The deeper failure mode is worse. Handed a void, a language model does what any prediction engine does: it fills the void with its priors. That is not malice, it is arithmetic. The prior says a crypto article published this week probably concerns a token, a launch, a raise, a partnership. So the completion arrives fluent, plausible, and connected to nothing on any chain.
The costs are concrete. Tracing a claim in late 2024 that a certain protocol had been "audited by" a well-known firm, I found the same sentence in fourteen articles, three newsletters, and one exchange listing page — and no audit report anywhere. Fourteen citations, one original error. Repetition is not corroboration, and models are, structurally, repetition engines. An industry that reads its own output as evidence has built a hall of mirrors and called it a research desk.
Labelling sits upstream of every conclusion. If a pipeline tags an asset a "Bitcoin Layer 2" because the project's own homepage says so, every downstream conclusion inherits that framing: it gets sized against Bitcoin, reasoned about in Bitcoin's vocabulary, marketed to Bitcoin's holders — while the actual code might be an EVM chain with a marketing department. Tags drawn from market copy rather than from code will always describe the pitch, never the machine. A label is an inference, and an inference has no business being laundered into an input.
I have seen the same category error inflate the other side of the market. Dashboards that sum active addresses across a dozen rollups and call the total "users" are counting one wallet a dozen times. That figure is real in the arithmetic sense and meaningless in every other sense. It is not growth. It is a photocopy.
None of this is an argument against automation. My own work depends on it. It is an argument about ordering. Verification has to be computed before synthesis, because synthesis is downstream of every error you have already made. A pipeline that checks its own inputs is boring, slow, and correct. A pipeline that skips that step is fast, fluent, and confidently wrong — which, in a market that prices information in seconds, is the same thing as dangerous.
What struck me most about the empty report was its shopping list. Before it would analyze anything, it wanted: a title, three to five information points with named sources, one sentence of core claim plus the author's stance, and at least one identified protocol or project. That is a genuinely defensible minimum, and it maps onto something I care about a great deal. Frameworks requiring real inputs cannot be run on vibes. The Howey test asks four concrete questions — money invested, a common enterprise, an expectation of profit, the efforts of a third party. You can argue about the answers forever. You cannot answer them without facts. Any analyst producing a securities assessment from a ticker and a thread has not applied a framework; they have decorated an opinion.
I keep returning to a habit I formed during those collapsed-protocol audits, and to something I owe the reader here. That final page in every report — the list of things I could not verify — was the only page that aged well. Investors hated it. Founders hated it. I kept writing it. I write under my own name, and I try to say plainly when I do not know, because anonymous, unattributed certainty is precisely the artifact this industry has too much of already.
Here is the uncomfortable part. The market does not reward the page that says "we do not know." In a bull market, the demand is for certainty, and certainty is what gets shared. A document that opens with "insufficient information" has no thread potential, no chart to screenshot, no quote to lift. So pipelines get tuned, quietly and incrementally, to never come back empty — the same way trading desks almost never publish "no position." The incentive gradient points one direction, and generated text follows gradients.
But the absence of a signal is itself a signal, and it is the one the entire downstream stack depends on. Every fabricated analysis inherits the original break and amplifies it, and by the time it reaches a reader it has acquired the texture of conviction. The most expensive mistakes in this market are not made by people who lacked data. They are made by people who had data that did not exist.
There is a larger version of this. Generative systems now produce text faster than verification systems can check it, and we have spent a decade scaling synthesis while scaling almost nothing on the provenance side. The bottleneck was never the writing. It was always the checking — and we automated the wrong half first.
What changes if every claim carries a lineage? If a reader can see, in one click, that a number came from a source that exists and was dated when it says it was? That is the direction the identity work I do now points: not policing speech, not ranking opinion, but making the difference between a finding and a fabrication auditable by anyone. The empty report was not a machine failing. It was a machine declining to lie. That should be the floor of this industry, not its ceiling. And quietly, over the next few years, it will become the standard every other report is measured against.
