Read the failure state first.
A crypto research pipeline was handed a source article and returned a nine-dimension analysis: technical architecture, token economics, market structure, ecosystem positioning, regulatory exposure, team and governance, risk matrix, narrative cycle, supply-chain transmission. Every substantive field read "insufficient information." The information point list was an empty array. The risk matrix had no rows to score. The investment value rating was five hollow stars out of five. And at the bottom, the pipeline graded its own meta-judgment — the judgment that it could not judge — as confidence: high.
That document is not a bug report. It is a control experiment that ran by accident, and it is the most useful thing I have read about crypto research this quarter. Because the only honest component in the entire output was the refusal, and in a normal production system the refusal is precisely what gets filtered out before a human ever sees it. What reaches the human is whatever fills the empty array.
Start with the cost curve, because the cost curve explains the behavior.
In 2017 I read roughly forty token whitepapers a month. Not skimming — reading, with a pen, counting the number of claims that survived the removal of adjectives. Two days per document. It was the only reason my fund survived that cycle, and it was unglamorous work that nobody paid for. The bottleneck then was never writing. It was verification. Producing a plausible forty-page token thesis cost a junior analyst three weeks. Producing a verified one cost six weeks and required someone who could read a Merkle proof without flinching.
That ratio inverted. Producing a plausible page now costs approximately nothing. Producing a verified page still costs days. When the cost of supply collapses and the cost of verification holds constant, you do not get more information. You get more claims per unit of information, and the ratio between the two becomes the only metric that matters.
Bear markets make this worse, not better. Demand for research rises — people want to know whether their assets are safe, which is the correct question — while willingness to pay for it falls, because unrealized losses dominate every budget line. The market responds the only way it can: volume. More coverage, thinner, faster. The pieces that get distributed are the ones that fit a template and carry a number in the headline.
There is a second shift layered on top. Since the spot ETF approvals, an enormous share of institutional crypto coverage has migrated away from protocol mechanics and into flow tables — creation, redemption, net inflows, basis. That is a legitimate subject. But it has trained a generation of readers to accept tabular confidence as a proxy for analytical rigor. A table of numbers looks like evidence. It is not evidence. It is formatting.
So we have a market where the supply of confident formatting is infinite, the supply of traceable claims is scarce, and the two are visually indistinguishable at a glance. That is a textbook adverse selection problem, and it resolves the way adverse selection always resolves: the low-quality good drives out the high-quality good unless someone pays for a signal.
Now look at what the empty shell actually proves.
The framework behind that report carried a constraint I have not seen enforced often enough: every conclusion must cite a specific information point from the source. Not "the source suggests" — an identifier. A pointer. When the pointer list came back empty, the framework had exactly one legal move, which was to refuse. It refused nine times and printed a data requirements list instead, prioritized P0 through P2, with an explicit note that two of those fields would unlock roughly seventy percent of the analytical surface.
That is an attestation requirement, and it is the correct architecture. It is also rare, because attestation requirements are expensive and they make pipelines look broken. A pipeline that refuses produces no artifact. A pipeline that fills produces a report with your logo on it. The incentive gradient points in one direction and it is not toward refusal.
The real failure mode is not hallucination. It is the gate. Everyone is worried about a model inventing a TVL figure. The more dangerous event is quieter: an empty input entering a downstream stage and being read as an absence of risk. One line in the diagnostic names it precisely — if a shell like this gets processed as "low risk, no anomalies," you produce systematic misjudgment, because missing evidence is not evidence of safety. Silence and safety look identical in a table and are opposites in reality. Every risk manager I have hired in the last four years has had to unlearn that equivalence, and the ones who could not unlearn it did not survive the 2022 restructuring.
I watched this exact confusion destroy balance sheets. Terra was not a data-poor system. It was data-rich and interpreted by people who read the absence of a red flag as a green one. The collapse was not a surprise to anyone who counted the claims in the Anchor documentation that could be traced to a cash flow. There were very few. The yield was the claim and the claim was the yield.
So build the test. Count claims, count citations, divide. I call it the provenance ratio, and it is the same instrument I used on whitepapers in 2017 with a pen instead of a script. Take any research document. Mark every assertion that would change a position if it were false — the TVL number, the unlock date, the auditor's name, the governance threshold, the identity of the lead investor. Now mark how many of those carry a pointer to something a third party could re-execute: an on-chain transaction, a contract address, a filing, a signed commit, a named source with a date. Divide. A document with forty position-relevant claims and zero traceable inputs scores zero. It is not low-quality analysis. It is not analysis.
Run that arithmetic on the report I read. The score is undefined, because the numerator is zero and the document says so out loud. That is the highest possible integrity outcome for a zero-data input — and it is the outcome that any content pipeline optimizing for shipped artifacts would have rejected.
Here is where the analogy to my own domain gets uncomfortable. I have argued for two years that the data availability layer is overbuilt — that the overwhelming majority of rollups do not generate enough data to justify dedicated DA, and that the market has priced a solution to a problem most chains do not have. I stand by that. But the research industry has the inverse pathology: it generates an enormous volume of claims and posts no data layer at all. A rollup that publishes a state root with no underlying data behind it is making an unverifiable assertion, and the correct response is to treat the data as unavailable and refuse the state transition. A report that publishes a conclusion with no information points behind it is structurally identical. The correct response is identical.
Nobody applies it, because there is no forced exit. A rollup that withholds data gets slashed or stalls. A report that withholds provenance gets distributed.
The instrumentation side is unglamorous and it works. Field-completion rate as a hard gate — if fewer than a threshold share of required fields resolve, the input gets rejected before it enters analysis, not annotated afterward. Citation density per thousand words, tracked as a shipping metric the same way you track gas per transaction. A mandatory timestamp field, because market commentary without a publication date cannot be assessed for whether it has already been priced. These are boring controls. Boring controls are what separate a system that reports from a system that performs.
The verifiable part of this market is the verification layer, and that is where the capital is going. In 2026 my research group published on machine-to-machine micropayments, and the thesis we kept returning to was not really about payments — it was about what autonomous agents will pay for. An agent executing a strategy does not need prose. It needs a signed, timestamped, re-executable fact with a price attached. That market is small today and will not stay small, for arithmetic reasons: the cost of generating a claim has collapsed, the cost of verifying one has not moved, and the spread between them is the entire business. Verification is expensive because it requires re-execution, and re-execution is the one thing you cannot fake at scale.
Every credible figure I have seen on this — and I have seen the internal decks, so treat the public numbers with the suspicion they deserve — points to an attestation layer in the billions within a cycle, most of it billed per call, most of it invisible to retail.
Follow the gas, not the hype. Ask who is paying. In the research stack, the writer is paid almost nothing, the distributor is paid a little, and the attestation provider is paid per query. That flow tells you what the market actually values, and it is not the article. It is the receipt.
The consensus framing is that AI will flood the information environment with synthetic misinformation. I think that is backwards, and the empty shell is the proof.
That report was not produced by a model attempting to deceive. It was produced by a pipeline whose success metric was output volume, and it refused anyway because someone had written a citation requirement into the specification. The threat model was never generation. Generation has been cheap since the printing press, and cheap since 2017, when a human with a template could produce forty pages of optimistic token prose in an afternoon. What changed is the unit economics of the flood, not its direction.
The actual blind spot is our treatment of uncertainty as failure. In risk work, "insufficient data" is a first-class finding — most of the time it is the finding. A system that cannot express it will express something else instead, and the something else will be a number. We have built a decade of dashboards with no field for unknown, so the unknown gets rendered as a zero, and a zero reads as safe. That is not a model problem. It is a schema problem, and it predates the models by a wide margin.
And note the cycle we are in. A bear market does not increase tolerance for ambiguity. It destroys it. Readers are down, they are scared, and they want a verdict. The pipeline that says "I cannot assess this" loses the click to the pipeline that says accumulate. Follow the incentive, not the technology.
The question to ask about any research you consume over the next twelve months is not whether a model wrote it. It is whether it can be falsified — whether there exists a version of the world in which the document is provably wrong, and whether the document tells you how to check. If it cannot fail, it is not a forecast. It is formatting with a logo.
Watch the attestation market, not the content market. Watch who pays per verification call, and watch what happens when the first agent refuses to settle against an unverified claim. That refusal is the product. Bets are cheap; exits are expensive.
And when a report hands you nine dimensions of confidence built on zero data points, read the empty array first.