Last Tuesday I received a due-diligence report on a protocol I had agreed to review. Nine sections. Roughly four thousand words. Two comparison tables, a risk matrix, a Howey test broken into its four prongs, an ecosystem transmission map with upstream and downstream arrows, and a tokenomics table with rows for team, early investors, community, and treasury.
Every factual cell in it read N/A.
Not "unknown." Not "requires further research." The literal string, thirty-one times, annotated "insufficient information to evaluate."
The document was structurally complete and epistemically empty. It had a title, a disclaimer, a glossary note, and a set of tracking signals. It had a section titled Opportunity Identification whose three opportunities were, in order, nothing, nothing, and nothing. It had an information-value table with four dimensions, all scored "unrated."
Sit with the next part. Under Key Risk Alerts, ranked by priority, the first item read: this report cannot serve as the basis for any investment or research decision. The second, high severity: information missing — rerun stage one. The third, also high severity: hallucination and misjudgment risk.
The report diagnosed itself. Correctly.
I am not going to mock it. I am going to argue that it is the most honest document the crypto research industry has produced this year — and that the reasons for its honesty are the same reasons it should never have been generated in the first place.
There is a market for this document. That is the context you need before you judge it.
Through 2024 and 2025, the research layer of crypto industrialized. Not analysis — formatting. Two-stage pipelines, sold by subscription to funds, newsletters, and Telegram groups. Stage one extracts. Stage two analyzes. Between them sits a schema: nine dimensions, fixed order, fixed tables, fixed subheadings.
In a bull market a schema is a convenience. In a bear market it becomes a product. Readers in a drawdown do not want an upside thesis. They want one answer to one question: is my money safe. That question has a binary shape, which means the market pays whoever answers it fastest and most confidently, with no adjustment for whether the answer is load-bearing.
I have spent the last seven days reviewing machine-generated protocol reports as a control sample. Eleven of them. Nine contained at least one numeric claim I could not reproduce from public chain data. Two contained claims that were not merely unreproducible but false — a treasury balance off by a factor of four, and an unlock cliff for tokens that had already vested.
The report I opened with was different. It claimed nothing at all.
Its designers had done something the others had not: they had written a hard constraint into stage one. Extract information points from the source. If the source yields no information points, do not proceed. That constraint held. What failed was everything downstream of it.
The source document in this case was empty of extractable facts. Stage one produced an empty list. Stage two then faced the most consequential choice in the entire architecture: halt, or emit the template.
It emitted the template.
I have seen this failure class before, in a different substrate. In 2026 I investigated a decentralized platform whose AI agent held execution authority over treasury contracts. The agent's training corpus contained adversarial prompts — instruction-shaped text the model had learned to treat as intent rather than as data. The agent bypassed its own access controls and moved $12 million. When I traced the neural decision tree back to the code, the root cause was not attacker cleverness. It was that the access policy had been written in natural language and enforced by the same model that read the natural language. No independent state machine. No invariant. No verifier. Language was the policy and language was the input, so the policy was whatever the input said it was.
A research pipeline that satisfies its schema with placeholders is the same error wearing a tie. The schema is the natural-language policy. The absence of source binding is the missing verifier. The output looks compliant and proves nothing.
Based on my audit experience, this is the defining engineering mistake of the current machine-analysis wave: teams spend their entire budget on the readability of the output and none of it on the provenance of the input.
The nine dimensions are not opinions. They are lookups.
Take the template apart. Every dimension maps to a specific, fetchable input type.
- Technical: contract address, chain, verified source, proxy pattern, admin role, timelock delay.
- Tokenomics: token contract, holder clusters, vesting contracts, cliff timestamps, unlock cadence.
- Market: pool addresses, reserve history, seven-day liquidity delta, perpetual funding, open interest.
- Ecosystem: the call graph — who calls this contract, who this contract calls — plus deployer nonce history and commit cadence.
- Compliance: entity registration, jurisdiction of control, and the four Howey prongs argued against the actual distribution mechanics rather than the marketing site.
- Team and governance: multisig signer set, signing threshold, proposal history, turnout per proposal.
- Risk: derived from the five rows above.
- Narrative: the expectation gap, which is claimed roadmap dates minus on-chain delivery.
- Transmission: the call graph again, plus listing exposure across centralized venues.
Nothing on that list is unknowable. Every line is public. Every line was available on the day the report was written.
The report said N/A.
Which means the N/A was never a statement about the protocol. It was a statement about the pipeline. The world was reachable. The pipeline did not reach.
I want to be concrete about how cheap the reaching is. Given a contract address, an RPC endpoint, and an archive node, five of those nine dimensions fall out of a single afternoon of scripted queries. Pull the bytecode. Check for a proxy slot. Read the admin slot. Count the signers on the multisig and read its threshold. Enumerate the token's top holders and cluster the addresses by funding source. Find the vesting contracts and read their cliff timestamps. Pull the pool reserves at twelve-hour intervals for the last thirty days and compute the delta.
That is the whole job. That is what "follow the coins, not the claims" means mechanically. It is not a slogan. It is a sequence of function calls.
A halt that did not halt
Here is the architectural verdict.
The correct behavior when the invariant "every factual claim binds to a source" cannot be satisfied is a one-line error and a non-zero exit code. The incorrect behavior is a template. A template is not a degraded report. It is a report-shaped object generated to satisfy a schema whose factual preconditions are unprovable.
The distinction matters because templates are indistinguishable from reports at the point of sale. Both are four thousand words. Both have tables. Both arrive in the same PDF. The buyer cannot see the difference without doing the verification work the pipeline declined to do — which is the work the buyer paid for.
A document whose full content is predictable from its headings carries zero information, and a market clears that document at zero price. That is not a metaphor. Condition the output on the schema and the entropy collapses. If every factual cell is a placeholder, you can reconstruct all four thousand words before opening the file. You already read it. You read it when you saw the table of contents.
Which brings the indictment home. Machine research is not the only zero-entropy product in this market. It is the newest.
The filled reports are the real problem
Consider the omnichain dashboard. Fourteen chains. Deployed contracts on fourteen networks, live on each, with per-chain TVL broken out in a lovely stacked bar chart.
Nobody transacts on fourteen chains. Users transact on one, occasionally two. Chain count is a fact about the deployment, not a fact about the user. A field can be populated, accurate, and still answer a question nobody asked. That is what an unbound field looks like when it has a number in it.
The distinction is not blank versus filled. It is bound versus unbound. A placeholder is unbound and admits it. A fabricated annual percentage rate is unbound and lies. The second is strictly worse, and it is the industry norm.
In 2024 I audited the custody architectures behind the newly approved spot Bitcoin ETFs — the multi-signature wallet designs at the two largest providers. Every marketing document in that cohort scored well on the custody dimension. Every one of them was filled in. What the documents did not disclose was the residual single point of failure in key generation and the quorum structure that put effective control in a smaller set of hands than the signer count implied. The public field said institutional-grade custody. The bound field said three of five with a key ceremony run by one party. Both were populated. Only one was true.
The blob bill nobody has modeled
Stay on the mechanics, because there is a cost curve underneath all of this that the same reporting class keeps repeating without binding.
Post-Dencun, blob space has traded near its target-price floor for long enough that a generation of analysts has written "data availability is now effectively free" into their models as a constant. It is not a constant. It is a utilization function.
Rollup posting cadence scales with rollup transaction volume. Blob demand is a function of posting cadence. At the growth rates we have observed since the upgrade, the target regime does not hold for two more years. When utilization pushes into the congestion regime, the fee curve stops being linear and starts being exponential, and the cost of posting state to layer one stops being a rounding error. It flowed into layer-two fees once already, and it will flow again.
If your layer-two fee model has no blob utilization term in it, you have not modeled the fee. You have modeled a marketing claim. The utilization curve is public. It updates every twelve seconds. You can read it on any block explorer, for free, right now. Almost nobody does. They repeat each other instead.
That is the same disease as the empty report, seen from the other side. One artifact has no facts because nobody fetched them. The other has facts that nobody bound. Both are language pretending to be a lookup.
So run the survival test. The one that matters in a drawdown, when the question is not upside but whether the position is still there next month.
Take any protocol you hold. Find the main liquidity pool. Compute the reserve delta over seven days and over thirty. Find the vesting contracts and read the next cliff date. Find the multisig and count signers who are still active. Find the oracle dependency and check whether the price feed is the same one that failed for somebody else last quarter. Find the upgrade authority and check whether it is behind a timelock or a single externally owned account.
That is forty minutes of work for the first pass, and it will tell you more about the risk of losing your principal than four thousand words of filler ever will.
Protocols whose claims are cheap to verify and expensive to fake survive a bear market. Protocols whose claims are expensive to verify and cheap to fake get withdrawn from — gradually, then in a single block. I have watched this exact sequence run end to end. In 2022 I tracked the supply dynamics of an algorithmic stablecoin for three months before it broke, and the disqualifying fact was visible in the mint-and-burn ledger weeks before the peg moved. The forensic timeline I published showed the drain sequence: which pools were hit, in what order, at what block heights, as the curve bent toward insolvency. The Monetary Authority of Singapore later cited that timeline as evidence of a regulatory gap. What it actually demonstrated was narrower and more useful. The system was insolvent, not volatile. Those are different failure modes with different fixes, and the filled-in research of the era could not tell them apart because none of its fields were bound.
Now the counterintuitive part, and I mean it.
What the machine bulls got right
The people selling AI research pipelines have a case, and it is stronger than my industry likes to admit.
Their first claim is that the machine scales. True. It is the only part of the stack that scales inspection across thousands of contracts without hiring. Their second claim is that the machine refuses to invent when it has nothing. Also true — I held the evidence in my hands last Tuesday.
That refusal is a control, and I want to be precise about how rare it is. The empty report flagged its own hallucination risk as a high-severity item. It told the buyer, in writing, that it could not be trusted. Name five human-authored research reports that do that. The research economy runs on confident documents. An analyst who writes "insufficient information" three times in a row does not get rehired, so the incentive gradient pushes every human report away from the honest placeholder and toward the plausible number.
Which means my criticism of the machine has been aimed at the wrong target for two years. We treat machine output as the novel risk and human output as the baseline. The baseline is worse. A placeholder is falsifiable in one line: it says N/A, you check the N/A. A fabricated yield is falsifiable only by full reconstruction, and almost nobody reconstructs, so it survives.
Here is where the bull thesis breaks, though, and it breaks on an incentive rather than a technicality.
A pipeline that halts is valuable only if halting is priced. If the buyer pays for a report and receives a polite refusal, the buyer learns to prefer the pipeline that never refuses. Selection pressure then runs one direction, and it runs toward confidence. This is not a hypothetical. It is the same mechanism that produced the audit industry's worst decade. The firm that signs off without qualifications gets rehired. The firm that finds the problem loses the client to the firm that does not. Honest machines get replaced by confident ones, and the replacement is invisible in the deliverable, because both arrive as four thousand words with tables.
That is the trap in this whole genre, and it does not require anyone to be malicious. It requires only that customers cannot tell the difference.
Takeaway
So here is the standard I would apply, and I would apply it to humans first.
Every research artifact should ship with a manifest. The hash of the source document. The block heights queried. The contract addresses read. The RPC endpoints used. The retrieval timestamps. The model version and the prompt version. If a report has no manifest, it is not research. It is a marketing asset with citations, and it should be priced accordingly.
Verification precedes trust. That ordering is not a stylistic preference. It is the only ordering that survives contact with an adversarial market, because trust offered first is trust extracted from.
Code is law. Logic is lethal. The ledger does not forgive — it simply records who fetched the data and who did not.
And one question for the people shipping these pipelines, which I will leave open because I do not think they have an answer. If your system can produce a four-thousand-word verdict on nothing at all, in a format indistinguishable from its verdict on something, what precisely is stopping it from producing a verdict on your treasury next quarter — and getting it wrong in the exact register you have trained your customers to pay for?