You think the dangerous research is the biased kind. It isn't. The dangerous research is the kind that is perfectly formatted and completely empty.
Last month I ran a nine-dimension audit pipeline against a live token: roughly $100 million fully diluted valuation, a landing page, a docs folder, a Discord moving fast, and three audit badges on the homepage. The pipeline returned 189 fields. Every single one of them said the same thing in different fonts โ information insufficient. Technical maturity: N/A. Supply structure by cohort: N/A. All five prongs of the Howey test: N/A. Six risk categories: N/A. Composite risk rating: unable to assess.
It was, structurally, the most rigorous document I have read this year. Tables. A dependency graph. A transmission map running from mining hardware through DeFi to end users. A glossary. A disclaimer. And it knew nothing. Not "knew little." Knew nothing, and said so forty-one separate times.
I sat with it for an hour. Then it clicked: the empty report was the most useful artifact I had seen in months โ not for what it said about that token, but for what it said about the machinery we have built around tokens.
Bull markets never have a capital problem. They have an attention problem. This cycle has produced more capital, more launches, and โ new this time โ more research than any cycle I have traded through since 2017. The research is no longer written only by humans. Agentic due diligence is the pitch of the season: feed a token into a pipeline, get back institutional-looking analysis in ninety seconds, subscribe for $499 a month and forward the PDF to your investment committee.
I have been on both ends of that transaction. In late 2017, in a Bangkok Telegram group, I manually audited fifteen ICO whitepapers and flagged eight of them as broken by opening the GitHub repositories. The check took eleven minutes per project: last commit date, whether the "smart contract engineer" had any commits at all, whether the repo was a fork with the README swapped. In 2020 I learned liquidity mining by losing fifteen percent to impermanent loss on my own capital, then taught two hundred developers in three workshops what that loss actually feels like. By 2022 I was certifying fintech professionals on AML, because the market had stopped rewarding anyone who could not answer a compliance question. In 2025 I started building curriculum for securing AI-agent wallets, which is how I ended up trusting โ and then distrusting โ automated analysis.
Here is the thing about due diligence that nobody sells: the value of a report is not its length, its structure, or its confidence. It is its information gain โ the number of things you know after reading it that you did not know before. By that metric, the 189-field report scored exactly zero, and it scored zero with perfect formatting, which is precisely what makes it dangerous. A blank page announces itself. A blank page wearing a risk matrix does not.
The table that grades the input
Look closely at the Howey section. Five rows โ money invested, common enterprise, expectation of profit, efforts of others, composite judgment โ and five N/A. Most readers will file that under "no data." That is the wrong read.
Howey is not a test you run on an asset in isolation. It is a test you run on a distribution and a set of promises. A token whose disclosure set is so thin that none of the five prongs can even be evaluated has not escaped securities analysis; it has failed it. The N/A is not a hole in the analysis โ it is the analysis. A launch engineered so that no legal framework can get purchase on it is a launch whose marketing was written by someone who understood exactly what to leave out.
I have watched this pattern since the ICO era. The whitepapers that scared me were never the ones with wild claims. They were the ones with no claims at all โ no team page, no token distribution, no roadmap, just a vision statement and a wallet address. Twenty-nine words of substance wrapped in forty pages of typography. The 189-field report is the same object, wearing better clothes.
A supply table with no supply
The tokenomics section is where the emptiness gets load-bearing. Team: N/A. Early investors: N/A. Community and liquidity: N/A. Treasury and ecosystem fund: N/A. Unlock schedule: N/A.
In eight years of reading these documents, unlock schedules are the single most predictive page in any token deck. Not the vision, not the partnerships, not the advisor headshots โ the vesting table. It tells you who is allowed to sell, when, and in what size relative to daily volume. When a pipeline cannot find one, the honest probability distribution is not "it doesn't exist." It is "it exists and has not been published," which is a directional signal, not an absence of one.
But here is where the pipeline betrayed itself. Total supply is not a secret. It is a return value. One call to the contract โ totalSupply() โ against any public RPC endpoint gives you the number in under a second. Whether the deployer retained mint authority is another call. Whether ownership was renounced, another. Code doesn't lie, but narratives do โ and this report was built entirely out of narratives. The machine read the docs folder and never touched the chain. That is not a data gap. That is a design choice, and it is the wrong one.
Alpha hidden in the noise. Everyone hunts for it in the docs, the Discord, the founder's posting history. The quietest source in the entire stack is the chain itself, and it answers in integers.
The machine didn't fail โ the schema did
I went back and read the pipeline's own diagnostics, because that is what I do. Three layers: retrieval, extraction, generation. Generation succeeded โ fluent, hedged, correctly scoped. Extraction succeeded โ it accurately determined that no extractable claim existed in the source. Retrieval returned zero bytes of primary material. The template held the shape while the content layer came back empty, and the system, being well-engineered, refused to crash.
That refusal is the bug. In software we learn this early: a schema is not a validator. JSON Schema will cheerfully accept an object where every field is null, provided you allowed null. Our research pipelines allow null. We built them so they would never crash, and they never crash, and they never tell us anything. The failure mode of 2026 is not hallucination. Hallucination at least makes a claim you can falsify. The failure mode is formal compliance with zero content โ a document that passes every structural check and carries no information, which the market then prices as if it carried some.
So I ran the check the pipeline skipped. Pulled the contract address from the docs, called totalSupply, read owner permissions, checked mint authority, pulled the top ten holders, and cross-referenced the deployer address against three previous launches. Four RPC calls and one block explorer tab. Ninety seconds. I learned more that was decision-relevant from ninety seconds of chain reads than from 189 formatted fields โ including the fact that the deployer had shipped two prior tokens with identical bytecode and different names.
That is the audit finding, and it generalizes: the more formatted a report is, the less you should trust its formatting as a proxy for work. Formatting is cheap. Bytecode is not.
The confession at the bottom
The report ended with a section listing what it would need in order to actually produce analysis: a title, at least one information point, the project or protocol name, and a timestamp. Four fields. It published its own dependency list, which is more honesty than most paid research desks manage in a year.
And that list is the price list. Four fields is what due diligence actually costs โ a name, a claim, a source, a date. Everything above it, the matrices and the transmission maps and the glossary and the disclaimer, is packaging. We have spent a decade building packaging and calling it analysis.
Consider what the transmission map actually accomplished. It diagrammed upstream infrastructure, midstream protocols, downstream users, and then wrote N/A into every node. The pipeline had a theory of how value moves through this industry and no evidence about how it moves through this token. That is a map of a country with no roads drawn on it.
This matters more this cycle than last, because the consumers have changed. In 2026 the buyer of a research report is frequently another agent. Agent-to-agent research markets are live, and they settle on structured output. Which means an empty report does not stay contained โ it enters a supply chain, gets embedded, becomes a retrieval source for the next pipeline downstream. A null propagates. I have watched enough systems to know that the cheapest way to corrupt a decision layer is not to feed it lies. It is to feed it confident formatting attached to nothing.
Here is the part that will annoy people who paid for the report. The empty document is more honest than ninety percent of the funded research published this quarter. The dangerous artifact in this market is not the one that says N/A forty-one times. It is the one that says strong accumulation with a total-addressable-market chart and not a single citation. Confident numbers are cheap to generate and expensive to falsify. An N/A is expensive to generate and cheap to falsify โ which is exactly backwards from what the market pays for. Trust is the new currency, and we are minting it against zero reserves, then marking it to model.
The reflex fix is worse than the disease. The obvious product response to an empty report is to fill the gaps with a language model โ infer the vesting schedule from a Medium post, estimate the team allocation from a screenshot, generate a plausible risk matrix. That converts an honest refusal into a confident fabrication, and it will ship, because customers prefer confidence and cannot audit it. A system that refuses to speculate is worth more than a system that speculates well, and nobody is pricing that difference.
I have made this mistake in the other direction. In 2021 I guided fifty Thai artists through minting contracts I was learning in real time, and I filled gaps in my own knowledge with confident answers because the room needed one. Two of those artists deployed contracts with royalty settings that never worked. Nothing catastrophic. But I remember the shape of it โ the moment where the format was ready and the substance was not, and I let the format speak. That is the same failure at industrial scale, with better typography and a subscription fee.
The next unlock in crypto research is not a better template. It is provenance. Every claim in a report should carry its source, a byte range, a timestamp, and a hash, so that a reader, human or agent, can distinguish a finding from a filler in one glance. Pipelines that cannot cite should refuse โ loudly, in the same font as everything else, and be priced accordingly. The question I keep returning to is not whether AI can write due diligence. It obviously can. The question is whether the person buying it can tell the difference between a report that found nothing and a report that looked for nothing.