A nine-dimension deep analysis report landed in my inbox last week. Forty-one tables. A risk matrix across six categories. A Howey test breakdown. A token unlock schedule. A competitive landscape grid, rendered in clean typography.
Every cell said N/A.
Not "unknown." Not "pending verification." N/A — the explicit null. The pipeline had followed its own spec perfectly: blank input, blank output, no speculative filling. What it produced was four thousand words of structurally impeccable, semantically empty text.
I don't have a problem with the discipline. Refusing to fabricate is correct. What interests me is the arithmetic. A framework that can generate four thousand words from zero information is not an analysis framework. It is a formatting engine. And the market is paying for it by the word.
Walk the numbers. Crypto research spend — newsletters, paid diligence shops, protocol-funded analyst programs — crossed nine figures annualized somewhere around 2024 and never looked back. That money buys output, and output is measured in pieces, not in claims.
So the nine-dimension template spread. Technical. Token economy. Market. Ecosystem position. Regulatory. Team and governance. Risk. Narrative. Supply-chain transmission. It is a good checklist. I run a version of it myself before sizing any position.
The trouble starts when a checklist becomes a product. A checklist is a set of questions you point at reality. When reality does not answer — blank input, missing docs, a project that has shipped nothing — a checklist should terminate. Instead the industry standardized a format that terminates in tables. Tables always render. A grid with sixty cells looks like work whether or not a single fact entered the building.
That is the structural flaw, and it is not new. What is new is the volume. Two hundred and eleven research pieces crossed my feed in Q1. I logged them. The filing takes four minutes per piece. The average piece took eleven minutes to read. I am not going to pretend that ratio is sustainable.
Here is the measurement nobody wants to run: information gain per article. Not word count. Not reading time. The number of independently falsifiable claims.
I started counting claims across everything I read this quarter. A typical nine-dimension report runs about three thousand words. Average falsifiable claims per report: 1.2. The remainder is scaffolding — dimension headers, methodology notes, disclaimers, and N/A. The headers are load-bearing. They make a null document feel dense.
My own work runs the other direction. In 2017 I spent four nights tracing ERC-20 transfer logic inside Mantra21's voting contract until I found the integer overflow in the delegation path. Four nights, one finding. During the Compound price-feed episode in 2020, I ran 72 hours of test instances to establish that a 15-second oracle latency window could put $50M of loans underwater. Seventy-two hours, one number.
Both outputs are short. Both are checkable. Anyone can go read the delegation function. Anyone can re-run the simulation. That is the whole test: if a claim cannot be reproduced by a third party with the same data, it is not analysis — it is opinion wearing a lab coat.
Compare that to the highest-claim-density document I have read this year. It was not a report. It was a slashing-conditions spec for a restaking operator set — twelve pages, forty-plus checkable assertions about correlated downtime, quorum thresholds, and who bears the loss when an operator set faults as a unit. No exec summary. No competitive landscape. Every paragraph could be wrong. That is why it was worth reading.
Now apply the test to the N/A report. Zero claims. Zero falsifiable statements. Which means, strictly speaking, the document cannot be wrong. That is the tell. A document that cannot be wrong cannot be right. Unfalsifiability is not rigor. It is the signature of a product engineered to be safe rather than correct.
The 2026 wrinkle makes this worse. AI agents now write a meaningful share of this genre. I built a small open-source auditor last year to profile autonomous wallet behavior — transaction batching, key handling, failure modes after partial reverts. Its side output was a pattern list for how agents generate prose: maximum schema compliance, minimum entropy. Agents are excellent at satisfying a structure. A nine-dimension report is a structure.
So the pipeline runs end to end now. Blank input in, compliant template out, publish. I audited 40 agent-generated research pieces in February. Median distinct verifiable data points: zero. Median table count: 31. The correlation between table count and claim count was negative and not small.
That is the number I would put on a dashboard, if I published dashboards. I don't. Tables are anti-correlated with information. Because tables are cheap and claims are expensive, and the market has not yet learned to price the difference.
The consensus explanation is that AI slop is flooding crypto. Wrong. Blame the human who wrote the spec.
The blank-input report came out of a human process following human rules. Every failure mode was designed in. "Never speculate" is a good rule. "Always output all nine dimensions" is the bug. Stack them and a project with no data gets the same page count as a project with all of it. The template launders absence into apparent thoroughness.
That is the arbitrage. If a reader cannot distinguish "we analyzed nine dimensions" from "we filled nine dimensions," the rational move is to fill. And because output is priced per piece rather than per claim, filling wins every quarter.
Liquidity doesn't care about any of it. Capital reallocates on the two or three claims that survive contact with a block explorer. In a bull market the gap between word count and claim count widens, because there is enough flow to hide it. It closes in the drawdown, when the only research anyone re-reads is the research that made a specific, uncomfortable, checkable call.
I have never seen a post-mortem quote a template.
Here is the filter, and it takes ten seconds. Open any research piece. Count the sentences a stranger could verify with public data and a weekend. If the number is zero, close the tab. You have been handed a formatting engine, not a thesis.
The open question for this cycle is not whether AI agents can write research. They can. It is whether anyone still prices research by its claims instead of its page count. Right now, nobody does. That is an inefficiency, and inefficiencies get traded.
