Nine Dimensions, Zero Data: The Silent Failure Inside Crypto's AI Research Layer

A due diligence report landed on my desk last week. Nine dimensions. Technical architecture. Token economics. Market structure. Ecosystem position. Regulatory exposure. Team and governance. Risk matrix. Narrative and expectations. Supply-chain transmission. Every cell read the same three words. Not available. Not "insufficient evidence." Not "inconclusive." Not available. The extraction layer upstream had returned an empty array, and the analysis layer downstream, being obedient, had produced a flawlessly formatted document describing nothing at all. The document ran to several thousand words. It had tables. It had a risk matrix with six rows, all blank. It had a disclaimer. It had, at the very bottom, a polite note suggesting someone re-run the first stage. What it did not have was a single fact. I have been editing crypto news for twenty-two years. I have read fabricated whitepapers, plagiarized audit reports, and "research" that was three screenshots and a price target. This was the first time I had read a document that was honest about being empty — and watched it get forwarded to an investment committee anyway. Here's the part that should worry us. In a sideways market, research volume goes up. That is not a coincidence. When price stops giving signals, everyone starts manufacturing their own. Desk analysts publish more. KOLs post more. And increasingly, the first draft of all of it is generated by an agent that scrapes a source, extracts facts, and hands a structured payload to a second model that writes the analysis. Two models, one JSON object between them. That handoff is where the whole thing lives or dies. The pipeline looks like this. Stage one reads the article and emits a schema: title, source, article type, domain tag, domain confidence, core thesis, and a list of information points. Stage two consumes that payload and fills nine analytical dimensions. Stage three publishes. When stage one returns an empty array for information points and nulls everywhere else, stage two has no legitimate move. It can refuse. It can degrade. Or it can hallucinate. The system I saw chose to refuse, and then chose to dress the refusal in a nine-dimension costume. I spent part of 2026 in a room with fifteen people drafting what became the Tokyo AI-Crypto Ethics Charter — developers, exchange compliance leads, two regulators, and me. The argument that consumed the most hours was not about what an AI agent should be allowed to trade. It was about what an AI agent should be allowed to say when it does not know. We did not resolve it. We wrote down that it needed resolving, which is what charters do. This report is that unresolved question, in production. Let me be precise about the failure, because the difference matters. There are two ways an automated research pipeline breaks. The first is degradation: three facts in, three dimensions partially filled, six marked insufficient. That is a usable artifact. A human can see the gaps and go fill them. The second is null propagation: zero facts in, and the schema still validates, because an empty array is valid JSON and a null string is a valid string field. Schema validation checks shape. It does not check emptiness. That is the entire bug, and it is boring, and it is everywhere. ⚠️ If your pipeline asserts on types but never on cardinality, you do not have a validation layer. You have a formatting layer. The second-order problem is worse. When a human reads "not available" in a risk matrix, two misreadings are available. The first is that the analyst looked and found nothing bad. The second — the one that actually happens in committee — is that the field is a placeholder someone will fill in later, so it gets skipped, and the report advances on the strength of its formatting. Neither reading is what the document says. The document says: unknown. In due diligence there is a conservative principle that the source document got right and that almost nobody applies to AI output: unknown risk defaults to high, not medium. An unpriced risk is not a neutral risk. It is the only category of risk you can be certain you are not being compensated for. I learned that the hard way in 2017, during the EOS airdrop verification blitz. We were auditing fifty thousand wallet addresses across Telegram groups, trying to separate real holders from sybil clusters, and publishing a live trust score. We moved fast. We broke the distribution story three days before the mainstream press. And for about six hours on day two, our dashboard showed a clean wallet because our scraper had silently timed out and returned an empty set, and an empty set renders identically to a clean one. Six hours. One wallet. It could have been the whole dashboard. Speed without an emptiness check is not speed. It is a coin flip with good typography. The same lesson showed up in 2020, during DeFi Summer, when Compound's interest rate curve spooked retail into a sell-off. I sat down with the cToken model and decoded it line by line, and then I ran three Twitter Spaces with community leads to explain it in plain language. The model was not complicated. The panic came from the gap between what the numbers said and what people could read. That gap is exactly where a confident, empty report does its damage. So let me put a number on it. Field fill rate. Count the populated fields, divide by total fields, publish the ratio next to the conclusion. Below sixty percent, you have a lead, not a finding. Above eighty, you have something a human can review. The document on my desk scored under ten. It was labeled a report. There is a third layer here, and it is the one I care most about as an editor: the three-tier separation of what a source says, what a reasonable reader can infer, and what someone is guessing. Explicit. Inferred. Speculative. Every published claim should be tagged. Most are not. The best crypto research desks do this instinctively in prose — "the filing states," "this implies," "our read is." The worst collapse all three into the same declarative voice, and the reader has no way to price the difference. An AI pipeline that does not carry those tags through its schema cannot produce them at the end. It will flatten everything into one confident register. That is not a model problem. That is a schema design problem, and it is fixable this quarter. Now the counter-intuitive part, and I want to be careful here. The empty report was the most honest artifact in the room. Every other document in that committee packet was fluent. Full sentences. Named comparables. A valuation range. And I would bet a meaningful amount that at least one of them was built on three thin facts and a strong prior, because that is what the incentive structure produces. Research is measured in pages shipped, not in claims survived. Analysts get promoted on coverage, not on the number of times they wrote "we don't know yet." A system that returns nine blanks is a smoke alarm going off. A system that returns nine confident paragraphs built on nothing is a fire that already burned down the building and is now sending a newsletter about it. We built the wrong detector. Everyone is instrumenting for hallucination — does the model invent a fact? — and almost nobody is instrumenting for null propagation, which is the failure mode where the model invents nothing and the pipeline ships anyway. Hallucination is loud. Null propagation is silent, and silence scales better. ⚠️ The dangerous report is not the one that says nothing. It is the one that says something plausible in a format that implies verification. There is a version of this that connects directly to the 2021 Azuki investigation I ran. Twenty female creators, twenty interviews, and a grant program that came out of it. None of that would have existed if I had written the piece from floor prices and a vibe. The human layer was the information gain. A model with no access to the human layer should say so, out loud, in the first paragraph. So here is what I am watching for over the next two quarters, and what I would ask any desk that hands you AI-assisted research. Ask for the fill rate. Ask which fields were empty at extraction. Ask whether unknown risk was defaulted to high or quietly rounded down to medium. Ask what the pipeline does when stage one returns an empty array — does it refuse, degrade, or write? The charter we drafted in Tokyo says transparency is a design requirement, not a disclosure. I still believe that. But a requirement only binds if someone checks. The next blowup in this market will not come from a bad token. It will come from a good-looking document. Which one is sitting in your committee packet right now?