The document that landed in my review queue at two in the morning was, by every formatting convention, a finished research report. It had nine sections. It had tables with header rows and confidence columns. It had a risk matrix, a token supply schedule, a Howey test breakdown, a transmission map running from upstream to downstream, and a disclaimer at the bottom longer than most project whitepapers. It also contained, by my count, zero facts.
Six of the nine fields returned the same string: insufficient information. The domain tag read unclassified. The list of information points โ the atomic factual claims on which every downstream judgment is supposed to rest โ was empty. And the section that asked for the projects involved in the analysis instructed the reader to identify them from the information points above, which was a small masterpiece of circularity, a pointer aimed at nothing.
The most honest line in the entire artifact was the one nobody had commissioned: a short declaration that the author would not fill the void with speculation, because in this industry a fabricated project name or a fabricated TVL figure converts, with terrible efficiency, into a wrong allocation decision. Everything else in the document was scaffolding around that one refusal.
I have read thousands of these reports. I have written more than a few. And I want to make an argument that will sound strange coming from someone who has spent twenty-three years insisting that analysis must be rigorous: this empty document is more dangerous than a wrong one. And the refusal buried inside it is the single most valuable thing in the stack.
Code is law, but conscience is the interpreter. The interpreter was the only part of this pipeline that worked.
Context: how the template became the product
To understand why an empty report is a systemic hazard rather than an embarrassing glitch, you have to understand what these nine-dimension frameworks were built to do.
Before 2022, crypto research was mostly narrative. A desk would publish a thesis, the thesis would rhyme with the price action, and the diligence lived in a partner's head. The collapses of that year โ the ones that took the better part of a cycle's credibility with them โ changed the procurement logic of every allocator I know. Family offices, mid-sized asset managers, and the compliance committees above them stopped asking for conviction and started asking for comparability. They wanted to look at four hundred assets and see the same nine columns in the same order.
So the template was born. Technical architecture. Token economics. Market structure. Ecosystem position. Regulatory posture. Team and governance. Risk surface. Narrative and expectation gap. Supply-chain transmission. Each dimension with sub-tables, each sub-table with fields, each field with a confidence marker. It is, genuinely, an improvement on vibes. It is also a machine that will happily produce a document indistinguishable in form from research while containing none of its substance.
The ingestion problem is mundane and old. A paywall blocks the crawler. A page renders its body through JavaScript and the fetcher receives an empty shell. A site serves a bot-detection interstitial. An input arrives with a headline and a navigation bar and nothing else. Any of these produce the same downstream signature: an article object whose fields are structurally valid and semantically hollow. Add two more failure modes that are harder to see โ a field-mapping error that writes the extracted facts into a key nobody reads, and a domain classifier that never fires โ and you have a pipeline that can lose an entire document without raising a single exception.
That last one is the tell. In a system whose default domain is blockchain and Web3, an output of unclassified is not a neutral observation. It is a diagnostic. It means the classifier did not run, or ran and returned nothing, and nobody upstream was watching for the difference between the two.
I have seen this exact pattern before, in a different costume. In 2017, at the height of the ICO fever, I audited the contract logic for a data-provenance startup whose founders wanted a mainnet launch in eleven days to catch a market window. I found five critical issues in the encryption layer that would have exposed user metadata. I wrote them up. I refused to sign. The founders explained, with the patience of people talking to a child, that the report could ship with the vulnerabilities listed as future work, that the market would not wait for a hardening sprint. I left instead. The lesson I carried out of that room was not about encryption. It was that in an audit document, a blank section is never a clean section โ and a section marked pending is a blank section wearing a tie.
What the pipeline in front of me did right was refuse. Most do not.
Core: the schema is the vulnerability
The failure here is not that the model hallucinated. The failure is one layer deeper, and it is a design failure I recognize from smart contracts.
A schema that permits the string insufficient information as a valid value in every single field is a schema that permits a complete, well-formed, fully compliant report containing zero information. The permission is the bug. Every field has an escape hatch, and the escape hatch is indistinguishable at the type level from a legitimate finding.
I learned to distrust this pattern the hard way, because the EVM taught it to everyone who audits Solidity. Consider ecrecover. When you hand it a malformed signature, it does not revert. It returns the zero address. If your contract does not explicitly guard against that zero address, then a failed signature verification silently becomes a successful authorization โ any caller can pass the check, because the check returned a value and the value looked like an address. The function did exactly what it promised. It verified nothing and returned null, and null was accepted as a party to the transaction.
Now consider ERC-20 balanceOf. Query an address that has never held the token and you receive zero. Query an address that held a million tokens and sold them all and you also receive zero. The interface cannot distinguish never existed from now empty. Both are zero, both are valid, both serialize identically into every dashboard that will ever read them.
The most dangerous failure mode in on-chain data is not a wrong number. It is a null that has been granted the same shape as a number.
That is precisely what happened in this report. N/A is not a finding. It is the absence of a finding, wearing the same clothes. And once it wears those clothes, the laundering begins.
The laundering of null into neutral
Follow the document downstream. A human analyst reads a risk matrix in which every row says insufficient information and, if they are experienced, feels a small cold spot โ because they know that unknown is not the same as benign. A scoring model reads the same matrix and does something much worse. It sees no red flags. It assigns a neutral score, because the parser found no high-severity tokens in a field that contained none, and the field contained none because it contained nothing.
This is where I will be precise, because precision is the only thing that helps. In risk terms, the correct price of an unknown is not zero and it is not the price of a known-low risk. It is higher than both. When I ran audit engagements, the severity ladder I insisted on had a rung that most firms omitted: not reviewed sat above reviewed, minor issues. A component nobody had opened was treated as more dangerous than a component that had been opened and found slightly wanting, because the second had bounded downside and the first had none. That ordering felt counterintuitive to founders every single time. It is also correct every single time.
The crypto industry has an unusually poor instinct here, and I think I know why. We have spent a decade teaching ourselves that on-chain data is truth. Public, immutable, verifiable. And that is true โ for data that exists. It says nothing about data that was never written. A contract with no oracle update has no price, but its last stored price remains, and a consumer that does not check the timestamp will trade on a number that is hours stale and perfectly formatted. The oracle did not lie. It simply stopped speaking, and the silence was interpreted as a quote.
Every serious oracle integration I have reviewed includes a heartbeat check for exactly this reason. You do not ask whether the price is plausible. You ask when it was last written, and if the answer is older than the heartbeat, you halt. The report in front of me had no heartbeat. Its fields were populated by a process that had stopped speaking, and nothing in the schema required anyone to check the timestamp.
The gate that was never built
The framework's own author proposed the correct remediation, and I want to credit it because it is the right instinct: a hard validation gate. If the information point list contains fewer than three atomic, verifiable factual claims, do not proceed to analysis. Reject, alert, retry. Do not degrade gracefully into a beautiful void.
I would go further, because three is a number someone chose and the principle matters more than the number. What the pipeline needs is not a threshold but a provenance trail. Every report should carry, as non-negotiable metadata, the fetch status of its source, the byte length of the raw response, whether the body was server-rendered or client-rendered, and the timestamp of extraction. When I built verification tooling for a staking governance whitepaper in 2024, working alongside three legal experts and two senior protocol developers, the single hardest argument I had to win was that our compliance framework needed to document what it had not examined. The lawyers wanted a clean matrix. I wanted an appendix of exclusions. We shipped with the appendix, and it is the reason two asset managers adopted the document as a reference guide โ not because it was exhaustive, but because a reader could tell precisely where its knowledge ended.
A framework that cannot express its own ignorance is not a framework. It is a formatting exercise.
The deeper version of this problem is structural, and it is the one I find genuinely alarming in 2026. The readers of these reports are increasingly not human. Autonomous agents now ingest research feeds, score assets, and route capital without a human in the loop between the document and the order. A human analyst reading insufficient information feels the cold spot. An agent reads a field, finds a string where a number was expected, coerces it to zero, and proceeds โ because agents do not have intuition, and no amount of model capability substitutes for a schema that never told the difference between a measured zero and an unmeasured one.
This is why I built Verifiable Humanhood the way I built it. The project exists to prove that a participant in a DAO is a distinct human being without exposing who they are, using zero-knowledge proofs so that the attestation carries the claim and nothing else. But the design constraint that mattered most was not privacy. It was that the system had to be able to say no. A proof system that can only return yes is not a proof system; it is a rubber stamp with a hash function attached. The same holds for research. An attestation that cannot express absence is not an attestation.
The same emptiness, at industry scale
I want to widen the frame, because the nine-dimension report is a small instance of a pattern this industry produces at every scale.
There are dozens of Layer 2 networks now, and the aggregate dashboards look magnificent. Total value locked across the rollup ecosystem, charted over four years, is a beautiful curve. What the curve does not show is that a substantial share of that liquidity is the same capital, bridged and re-bridged, counted once per chain. The same tens of thousands of addresses appear in the unique-user column of nine networks simultaneously. The dashboards are not lying. They are reporting a real number whose meaning has been quietly severed from the thing readers assume it measures. Formatting is not liquidity. A full table is not a full market.
The same critique applies to my own tribe, and I will not exempt it. Decentralization advocates โ myself included, for years โ demanded transparency and then accepted dashboards as its fulfillment. We conflated on-chain with true. On-chain means recorded. Those are different words with different guarantees, and the gap between them is exactly the space in which an empty template can pass for a finished report.
There is a quieter version of this in every audit I have ever done. The checklist comes back complete. Every box ticked, every section filled, every row populated with an answer. And the answer in eleven of those rows is that the question was not applicable, or the component was out of scope, or the reviewer did not have access. The report is full. The knowledge is not.
Solitude is the only auditor that never sleeps, and solitude is also the only auditor that will tell you plainly which questions it could not answer. A committee will not. A committee fills the box.
Contrarian: the refusal came too late, and that is the real finding
It would be comfortable to end by praising the pipeline for refusing to speculate. I am not going to do that, because the praise would be a distraction from the more useful observation.
The refusal arrived only because the input was totally empty. Total emptiness is easy. A completely blank document triggers every alarm anyone ever built, including the crude one that says: six of nine fields are null, something is wrong. The dangerous case is partial emptiness. A report with seven dimensions fully populated and two marked insufficient information passes human review without friction, because it looks like thorough work with appropriate humility about the edges. It is not. It is a document whose author has learned exactly which two sections nobody checks.
And I want to name the incentive honestly, because moralizing about it accomplishes nothing. The market rewards coverage. The analyst who ships a filled template gets distribution, a byline, and a client renewal. The analyst who returns null gets replaced by someone who will not. This is not a character flaw in any individual. It is a gradient, and gradients move people. Every research desk in this industry is standing on a slope that pushes toward fabrication, and the only thing that holds position is a schema with a hard gate and a client base that has been educated to accept I do not know as a deliverable rather than a failure.
I learned how hard that education is during the three months I disappeared in 2022, after watching projects I had defended fail not on technical grounds but on human ones โ a small number of people, in a small number of rooms, choosing to fill the box. When I came back, my writing had changed. I stopped making predictions. I started documenting the conditions under which a claim could be falsified. Some readers left. The ones who stayed were the ones worth having.
The loudest voice is rarely the most aligned. The loudest voice is usually the one that never returns null.
Takeaway: design for nullability
So here is what I would build, and what I would demand of anyone whose research I read.
Nullability as a first-class primitive. Not an escape hatch bolted onto a schema, but a distinct type with its own shape, its own rendering, and its own consequences โ something an agent cannot coerce to zero because zero is a different type. Signed nulls: an attestation that a question was asked and could not be answered, carrying the provenance of the attempt. Refusal records preserved as assets rather than discarded as noise, because the map of what a system could not see is the most valuable document it will ever produce.
And a gate. Hard, unglamorous, and upstream of everything else: fewer than three verifiable facts, no report. Alert, retry, escalate. The most expensive line of code in any research pipeline is the one that stops it.
The question I keep returning to is not whether our analysis is accurate. It is whether our analysis can be wrong out loud. A system that cannot say I do not know will eventually say something else, and it will say it in a table, with a confidence column, at two in the morning, to someone who is tired and needs a number.