Last week I opened an analysis report that a research desk had paid good money to generate. It was for a mid-cap DeFi protocol that had just closed a nine-figure round. The report came back formatted, timestamped, and utterly clean. Every field was null. Technical structure: insufficient information. Token economics: insufficient information. Team and governance: insufficient information. Nine analytical dimensions, nine abstentions, and one line of apology at the bottom that said, more or less, the input was empty, so we refused to guess.
Nobody on the desk flagged it. Nobody sent it back. It sat in the shared drive for four days like an unpaid parking ticket.
That silence is the story. In a bull market where every project is fighting for narrative, the most honest document I read this month was a confession of ignorance β and nearly everyone treated it as noise.
Let me back up, because the empty report is not really about that one protocol. It is about the machinery that produced it, and about what the machinery is being asked to do in this cycle.
For most of my eleven years watching this industry, research was a craft practiced by humans with too much time and too little sleep. You read the whitepaper, you joined the Discord, you watched the order book for weeks, and you argued with strangers until the argument produced a thesis. It was slow, and it was messy, and it worked because the slowness was the point β the friction was where judgment lived.
Then we automated it. By 2025, almost every serious desk I know had wired up some version of an extraction pipeline: a model that reads a governance forum, a whitepaper, an on-chain ledger, a founder's X posts, and pulls structured facts into a table. The good ones run in two stages. Stage one extracts β names, numbers, unlock schedules, audit references, jurisdiction. Stage two interprets β it weighs those facts against a framework and produces a judgment.
Two stages is a deliberate design. Extraction is supposed to be boring and verifiable. Interpretation is supposed to be where opinion lives. The wall between them is the only thing keeping the output honest.
The report I found had a stage one that returned nothing. And that is where the interesting part begins, because a system that receives nothing has two choices, and the choice it makes tells you everything about who built it.
When a two-stage pipeline gets an empty stage one, it can lie or it can abstain.
The liar is more common than anyone admits. Ask a model to summarize an empty document and it will summarize the emptiness. It will invent a team, infer a tokenomics model from the category average, describe a roadmap that sounds exactly like every other roadmap, and hand you a page that reads like analysis. It is fluent. It is confident. It is fiction wearing a suit. And because the fiction is calibrated to the category β correct vocabulary, correct cadence, correct conclusions β it passes review. The desk I mentioned did not notice the empty report precisely because its earlier, hallucinated siblings looked normal. We have trained ourselves to accept the shape of analysis as a substitute for its substance.
The abstainer is rarer, and it is rarer for a reason. Abstention is expensive. It produces a document nobody wants, describes a decision nobody made, and β most importantly β it cannot be invoiced as insight. If you run a research desk, an empty report is a cost center. A confident report is a product. The market pays for the second one every single time.
So the pipeline I found was unusual. It carried a rule, apparently written in advance: if the source fields are empty, do not infer. It flagged nine missing categories, ranked them by priority β the information list, the core thesis, project identity, and domain tag as critical; source quality and time sensitivity as important; title and provenance as useful β and then it stopped. It explicitly refused to speculate. It even listed what it would need to proceed: at least three to five concrete information points, a project name, a one-sentence thesis.
I have seen that discipline in maybe a handful of systems. I have seen the opposite in almost all of them.
I learned this lesson in the summer of 2020, on a Discord server in Vienna, in the middle of a rebasing protocol called Ampleforth. Five thousand daily active users, most of them terrified. The elastic supply mechanism was elegant on paper and incomprehensible in practice, and every time the supply rebased, the support channel flooded with the same question in twelve phrasings: did I just lose money?
The answer, mathematically, was no. The answer, emotionally, was that they had no idea what they held, and that uncertainty was doing real damage. I stopped explaining rebasing math and started drawing pictures. Support tickets dropped by 40 percent. Nothing about the protocol changed. What changed was that someone translated the mechanism into human terms, and the translation restored a feeling of control.
I think about that every time I look at a research pipeline. A fact extracted and left untranslated is not yet knowledge, the same way a rebasing formula left unexplained is not yet trust. The extraction stage gives us facts. The interpretation stage is where we owe the reader the translation β and it is precisely the stage where hallucination does the most damage, because a translated lie is more persuasive than a raw one.
A year later, in 2021, I ran a grassroots study on the meme economy β a hundred and fifty interviews with holders and creators β trying to understand why absurdity was generating value. The finding that stayed with me was that narratives precede utility in early adoption. People did not buy because the fundamentals were strong. They bought because they recognized themselves in the story.
Hold that next to the hallucination problem, and the tension becomes obvious. If narrative precedes utility, then in the early life of any project the data genuinely is thin β sometimes so thin that an honest stage one comes back empty. The empty report is not a bug in the pipeline. It is an accurate description of an early-stage asset, and the industry cannot metabolize accuracy about early-stage assets, because accuracy is boring and boring does not raise a fund.
By 2022, after the collapse that shattered the cycle, I stopped writing research the way I used to. I ran a support circle in Vienna for junior analysts who were falling apart. Ten small sessions, fifty people, a lot of silence. What that year taught me was that resilience is communal, not individual β and that the most useful thing a report can do in a downturn is tell you the truth slowly.
Something similar is happening now, at the level of the machines. This year I published work on what I called the Empathy Algorithm, looking at how AI-driven organizations manage community sentiment. The pattern was consistent: agents that could execute governance but could not narrate it lost their communities within months. Loyalty did not attach to efficiency. It attached to context.
Now flip that finding around, because here is what unsettled me. An agent that invents a narrative when the data is missing is doing exactly what a persuasive human does when the facts are thin. The difference is scale. A human can hallucinate one thesis a day. A pipeline can hallucinate ten thousand an hour, all of them formatted, all of them confident, all of them passing the same review that let the empty report sit unnoticed in a shared drive for four days.
We built the extraction layer to be verifiable and the interpretation layer to be opinionated. What we did not build, until recently, was the third layer: the one that can say no.
I have started calling it the abstention boundary. It is not a model feature; it is a governance decision. Someone has to decide, in advance, what an acceptable minimum of information looks like, and what the system does when that minimum is not met. Most teams never make that decision, which means the default answer is the liar's answer: proceed.
The pipelines that refuse to guess are the ones I trust most, and they are the ones most likely to be defunded. That is the whole problem in one sentence.
Here is the part where I have to argue with myself, because the comfortable version of this essay writes itself: empty reports are honest, hallucinations are dangerous, abstain more often, the end.
The contrarian reading is harder. An empty report is not evidence of a data shortage. It is evidence of a parsing failure. The information was usually there β in a forum thread, a group chat, a chain explorer, an obscure governance post in a language nobody on the desk reads. The pipeline returned empty not because the project is opaque but because the pipeline could not see. We have spent two years building systems that ingest English and structured data, and we call everything else insufficient.
That reframing matters, because the industry then does something worse than hallucinate: it conflates no data with no value. A project with a thin paper trail and no English-language presence gets filtered out of the deal flow β and that filter is exactly where the asymmetric opportunities live. I watched this in the meme cycle. The projects with the cleanest documentation underperformed the projects with the messiest communities. Documentation is not the same as adoption.
So the empty report is neither a triumph of discipline nor a scandal. It is a mirror. It shows you the boundary of what your system can perceive. The question is not whether to abstain. The question is what you do with the abstention: file it away, or read it as a signal that your eyes are not good enough yet.
The next narrative in crypto research will not be about bigger models. It will be about provenance and permission β systems that can document where a fact came from, and systems that are allowed to say no.
Sometime in the next year, someone will ship an abstention score, a single number that tells you how much a report actually knows, and desks will start pricing it into their decisions. I hope they do. Because in a market where everyone is paying for confidence, the most valuable thing you can sell is a reliable way to tell when confidence is fake.
So here is the question I keep circling. If we cannot teach our machines to admit an empty page, what makes us think we can teach them to earn our trust β and the story isn't in the token, it's in the trust, after all?
