In late March, a member of our Tallinn reading group forwarded me a due diligence report on a freshly funded Layer 2. Forty pages. Nine analytical dimensions. A color-coded risk matrix, a four-factor Howey test table. Every field rendered in the same confident typography, and every conclusion marked insufficient information. The pipeline had produced a complete document containing exactly zero findings.
My first reaction was relief, because I have spent years arguing that a system which cannot answer should say so. My second reaction was colder. Nothing in that document's visual grammar told you it was empty. If the recipient had scrolled past page one, they would have assumed diligence had occurred, and the formatting would have backed them up.
That is why I want to write about it. The same pipeline, pointed at a source containing three sentences, would certainly have returned four thousand words of confident prose. The empty report was the exception. It should be the rule.
The architecture behind that document is now standard across crypto. It runs in two stages. The first extracts discrete information points from source material — a whitepaper, a governance forum thread, an on-chain dashboard, a founder's podcast appearance. The second runs those points through a fixed analytical schema: technology, token economics, market structure, ecosystem position, regulatory exposure, team and governance, risk, narrative, supply-chain transmission. Funds use it. Grant committees use it. Listing desks use it. Increasingly, so does the research tab of a project's own website.
This is a rational response to volume. I ran workshops for two thousand Estonians during DeFi summer in 2020, and the hardest thing to teach was triage — how to decide, in fifteen minutes, whether a protocol deserved another fifteen. Automating triage is not laziness. It is the only way to keep up.
But this cycle has inverted the economics. Deal flow has multiplied while diligence capacity has not. Dozens of Layer 2 networks now compete for the same finite pool of users, and each arrives with a nine-figure raise. A $100M announcement does not create more analysts. It creates more pressure to produce a document. That pressure is where the pipeline learns its habits.
The failure mode is not the model. It is the schema. Somewhere in that pipeline is a configuration file listing required fields, and required means required. When a system cannot fill a required field, it has exactly two options: fail, or fabricate. Production pipelines almost never fail, because failing produces nothing, and nothing cannot be shown to a committee.
Language models are tuned on human preference data that rewards helpfulness over abstention. Ask a well-tuned model to describe the token distribution of a project whose tokenomics page does not exist, and it will describe the token distribution. "The document does not contain this" reads as unhelpful. Unhelpful loses. The reward signal does not distinguish between a description drawn from a source and a description drawn from the shape of the question. Both are fluent.
Every required field is an incentive to invent, and every invention inherits the authority of the template around it.
Code binds, but people break or build — and here the binding code is the schema, while the people are the ones who wrote the schema, tuned the reward model, and clicked approve on the batch job. Nobody in that chain experienced a decision to lie. They experienced a decision to ship.
The second-order damage is where it gets expensive. A generated report does not stay a report. It becomes an input to the next system — a grant committee's scoring rubric, a listing desk's checklist, a newsletter's summary, a treasury's allocation memo. Each hop launders uncertainty into confidence. The half-life of a fabricated claim in crypto is extraordinarily long, because nobody re-reads the input. They read the output, and the output was well typeset.
What would an honest version look like? I have been pushing a specific design in the working groups I sit in. Every report carries an input manifest: content hashes of the source material, retrieval timestamps, and a coverage ratio — the share of schema fields for which a primary source actually exists. A report with a zero coverage ratio should be a signed null, published as a null, and counted as work. Abstention has to become a first-class output, or it will remain a career risk.
This is the same problem my colleagues and I attacked in 2025 through the Human-Centric AI Alliance, where fifteen researchers spent a year reading twenty papers and ten pilot deployments looking for a way to keep machine-mediated judgment anchored to something verifiable. We found a narrower answer than we wanted, and a more useful one. Zero-knowledge proofs let you verify a computation without revealing its inputs. We need something adjacent but different: proof of what was read. Not proof of comprehension. Proof that material existed, that it was retrieved at a stated time, and that it is the material that was analyzed. Input attestation is unglamorous. It is also the entire difference between a research pipeline and a rumor mill.
The human layer matters more than the technical one. In twenty workshops across 2020 and 2021, the most valuable sentence I ever said was: I don't know, and here is what would have to be true for us to find out. Almost nobody in those rooms had been told that sentence was available to them. The pipeline has not been told either.
Culture eats blockchain for breakfast. A team that prizes the appearance of rigor will purchase, or train, systems that produce the appearance of rigor. The model is downstream of the culture. It is a mirror with a temperature setting.
One more detail about that empty report. Its footer said automated. It was not automated. A human selected the nine dimensions. A human decided which fields were mandatory. A human scheduled the batch job. Three people, one schema, and a cron expression — which is, if you squint, the same governance structure as most organizations that describe themselves as leaderless. The keys to the upgrade sat with a small group, exactly where they always sit. Code binds, but people break or build — including the code that decides what counts as a finding.
Here is where I expect pushback, and I have argued this at length with people I respect. That empty report was the system working correctly. A validity gate caught a null input and refused to hallucinate. If we want fewer fabricated analyses, the sane response is to ask for more documents that say nothing — publicly and verifiably.
But blaming the model is the industry's favorite evasion. The demand side built this. We ask for a deliverable, not an answer. Nine dimensions are a format, not an epistemology. A framework is a promise about where to look. It is never evidence that anyone looked.
And I will indict my own archive. In 2017 I screened more than fifty ICO whitepapers and judged twelve viable. The other thirty-eight did not receive reports. They received private rejections that were never published. Nobody publishes the N/A files. The result is that the public corpus of crypto research is survivorship-biased toward projects that happened to have enough material to analyze, while the honest nulls rot in a folder. We only ever see the filled tables. A system that only ever ships filled tables will learn to fill them.
So: when the next glowing report reaches you — and in this market it will — ask for the input manifest before you ask for the score. Ask what the extractor actually read, and when, and whether the pipeline has standing permission to return null. Trust is the only currency that matters, and it is not issued by typography. We are building the future, together — but only the parts we can prove we read first.