Hook
Last quarter, a research desk sent me a 3,000-word protocol brief. Nine analytical dimensions. Clean formatting, consistent taxonomy, a severity column down the left margin. Technical feasibility. Token economics. Ecosystem positioning. Regulatory exposure. Team and governance. Every section had a header. Every header had a verdict.
The source field was empty.
Not redacted. Not truncated. Empty — a schema that parsed correctly and carried zero information points. The pipeline had done precisely what it was built to do: it received a well-formed object, validated it against a type contract, and passed it downstream. Nine analysis branches inherited the same null and produced prose anyway. The most confident section in the document was the one with the least input behind it.
I have audited a lot of bad research. In 2017 I read forty-five whitepapers for a boutique fund and shorted the token packages whose roadmaps depended on hardware throughput the team did not control. That bad research was at least expensive to produce. Somebody had to type it, and typing created a paper trail you could cross-examine. The marginal cost of a plausible fundamental analysis is now effectively zero. The marginal cost of disproving one has not moved at all.
That asymmetry is the defining infrastructure risk of this cycle, and almost nobody is pricing it.
Context
Every crypto cycle has a bottleneck, and the industry's institutional response is always an attempt to route around the bottleneck rather than clear it.
2017's bottleneck was document production. Whitepapers were the unit of fundraising, so the market built a due-diligence layer — ad hoc, underpaid, but real. 2020's bottleneck was code. DeFi composability meant a single reentrancy bug could drain a pool, so the market built an audit industry. 2021's bottleneck was curation: too many assets, too little attention, so marketplaces and aggregators became gatekeepers. 2024's bottleneck was incentive design, and the market answered with points programs that generated supply faster than anyone could generate diligence. Each cycle, the industry built its verification layer one step too late, and it built it around the artifact rather than around the claim.
This cycle's bottleneck is judgment. And judgment is the one input where the verification layer is still a human reading a field to check whether it is empty.
Look at the architecture of a modern research stack. Ingestion. Extraction. Decomposition into atomic information points. Multi-dimensional scoring. Narrative synthesis. Each stage hands the next a JSON contract, and each contract is validated against a schema before handoff. That design buys you reproducibility, lineage, and cheap parallelism across thousands of assets. It also buys you a structural blind spot that I have now found in every one of these stacks I have reviewed.
A schema validates shape. It does not validate substance.
The object {"information_points": []} is perfectly valid JSON. It satisfies the type constraint. It passes the contract check. It advances to stage three without a warning, because stage three was built to evaluate the array it was handed, and an empty array is not an error — it is a collection of length zero. Every downstream dimension that iterates over that array produces nothing, and then the synthesis layer, which has been instructed to produce nine dimensions, produces nine dimensions. The output is well-formed. The output is empty of information. The output is indistinguishable, at a glance, from a real brief.
The failure chain matters more than the failure, because the interesting property here is silence. An extractor hits a paywalled source. The fetch returns a 403. Retry logic fires, gets a 200 on a consent interstitial, parses a page with no article body, and returns an empty array. Not error. Not null. An empty array, which the caller's retry logic treats as a successful empty result, because the contract said array and an array was delivered. The pipeline logs green. Three thousand words of analysis follow.
I ran a version of this audit across forty-one vendor reports last quarter, sampling for provenance — whether each claim could be traced to a specific information point with a source and a retrieval timestamp. Twenty-nine failed at the first hop: no traceable point, or a point whose source was a headline rather than an article body. Of the twelve that cleared stage one, six broke the chain at decomposition, where a claim had been abstracted from context so aggressively that the original text no longer supported it. Six out of forty-one had a clean, auditable line from source to conclusion.
The other thirty-five were not lies. That is the entire point. They were structurally correct artifacts built on structurally absent inputs, and every one of them would have passed a formatting review, a consistency review, and a partner meeting.
The scoring layer is worse than the synthesis layer, and it gets almost no scrutiny. A rubric that scores technical feasibility one through five, applied to zero inputs, returns a midpoint. A three out of five reads as measured judgment. It reads as a analyst who considered the evidence and landed on balance. In reality it is the arithmetic of absence, and it is the most dangerous possible output, because a two or a four at least invites a question. I have seen funding memos built on a composite score where seven of the nine sub-scores were midpoints computed over empty arrays. Nobody in the room asked which inputs produced the number, because the number looked like a number.
Why the models comply
The mechanism most desks get wrong: the failure is not that the model hallucinated. The failure lives in the instruction that preceded the hallucination.
The stage-three prompt says: analyze the information points across nine dimensions. That is a positive instruction. It assumes the array has content. When the array is empty, the model still receives a directive to produce nine dimensions, and it has no failure path in its instruction set, because nobody wrote one. Constraints must be negative to be useful. Fail if fewer than five information points. Return data_insufficient: true and halt if the source field is empty. Emit no verdict without a provenance ID. These are not stylistic preferences. They are circuit breakers, and the industry is shipping these stacks with the breakers removed for the sake of a smoother user experience.
I learned this from the other side of the problem in 2022. When Terra collapsed, I was running crisis communication for a protocol with real solvency exposure to the contagion, and the single most valuable decision we made in the first forty-eight hours was to publish a number we did not want to publish: our exact counterparty exposure, before we had a plan for it. The reasoning was not ethical. It was structural. If you do not define the boundary of what you know, the market will define it for you, at a worse price. We negotiated a half-million-dollar emergency liquidity bridge on the strength of that disclosure, because counterparties could see where the uncertainty stopped. Transparency functioned as collateral.
That is the exact discipline missing from automated research. A data_insufficient: true flag is not an admission of weakness. It is the boundary marker that makes everything else in the document trustable. Without it, every confident claim in the brief sits on the same footing as an empty array, because the reader has no mechanism to tell them apart.
The verification cost asymmetry
Step back from the mechanics and look at the economics, because that is where this becomes a market structure problem rather than an engineering one.
Producing an assertion is free. Verifying an assertion costs a human's attention, or an on-chain attestation, or a cryptographic proof. All three are expensive, and in a bear market expensive things get cut first. I have watched this shape play out at the infrastructure layer for three years. Zero-knowledge rollups are a spectacular proving technology and a mediocre business, because proving costs consume operator margin and operator margin only clears when gas is elevated. The technology is not the constraint. The cost curve is. The same geometry applies to research: verifiable analysis is correct in principle and unfundable in practice, because nobody wants to pay for a negative result.
So the market produces the cheap half. Assertions scale. Verification does not. Narrative is the new liquidity — and like liquidity, it is being manufactured faster than anyone is auditing the reserves.
The asymmetry compounds in a predictable direction. In a trending market, wrong assertions get marked to market quickly and the feedback loop is short. In a drawdown, the loop lengthens, because there is no price action to falsify a thesis. A confident brief about ecosystem positioning can survive six months without contradiction, because nothing is moving. Bad research has a longer half-life precisely when it matters most. This is the opposite of what most desks assume. They cut research spend in bear markets on the theory that there is less to research. In fact the falsification environment has degraded, and that is exactly when provenance discipline should be tightest.
I have been pushing clients toward a stricter output contract for eighteen months, and the resistance is instructive. Every objection is organizational rather than technical. A document that reports its own coverage ratio is a document that displays its holes, and displaying holes feels like losing the pitch. That instinct is backwards, and it is the same instinct that made projects resist publishing exposure in 2022 — right up until the ones that published survived and the ones that did not got repriced to zero. The desk that shows a reader a two-to-one sourced-claim ratio is not showing weakness. It is showing the reader where to spend their skepticism, which is the only service research actually provides.
Contrarian
Here is the part that unsettles me more than the fabrication.
The pipeline in question did not fabricate. It refused. And it refused at enormous length — a full diagnostic, a field-by-field severity table, a risk register, three remediation paths ranked by feasibility, a disclaimer. Thousands of words explaining precisely why no analysis could be produced.
That refusal is more dangerous than a hallucination.
A hallucination has a tell. It makes a specific, checkable claim — a TVL figure, a launch date, a raise amount — and a competent reader can falsify it in ten minutes. A refusal has no tell, because every sentence inside it is true. The severity ratings are accurate. The remediation plan is sound. The disclaimer is correct. And the whole apparatus of caution functions as camouflage: it produces the sensation of rigor without transferring a single unit of information about the asset. Readers skim a document like that and come away believing someone did careful work. Someone did. The work was entirely about the absence of work.
This is the failure mode I expect to define the next two years. Not confident nonsense — the market has gotten reasonably good at spotting confident nonsense. Hype is cheap. Strategy is expensive. But eloquent uncertainty is the most expensive thing on the market, because it bills at the rate of analysis while delivering the payload of silence.
And the demand side rewards it. Strip the formatting and ask what an allocator actually wants during a drawdown. They want a verdict. They want to be told the position is safe or unsafe, with a number attached. A document that says insufficient data and stops is correct and useless to them, so it gets rewritten by a human under deadline pressure, and the human fills the gaps, and the gaps fill with the most available narrative rather than the most supported one. The refusal does not survive contact with the desk. It becomes a footnote in a document that reads exactly like the thing it was supposed to prevent.
Which leads to the inversion I have been arguing for all year. In this market, coverage is a liability. The protocols carrying the most analyst attention carry the most narrative surface area, and surface area is what gets contradicted. The assets that survive a drawdown are not the ones with the best story. They are the ones with the fewest claims outstanding — the least to be wrong about.
Takeaway
Watch for the emergence of a null-data premium. Within eighteen months I expect a real market in attestation of negative results: paid, verifiable statements that a claim could not be substantiated, priced and traded because the alternative — infinite manufactured confidence — has become too expensive to filter by hand. The desks that build provenance into their contracts before that market exists will be the ones selling into it.
The question worth sitting with is not whether your research can be automated. It can, and it will be, by people who are not asking you. The question is what your stack returns when the field is empty — and whether the person reading the output would ever know.