Nine sections. Forty-one tables. Zero facts.
On a Tuesday morning in late October, a research pipeline I had been stress-testing returned a fourteen-page document. The schema had rendered perfectly โ technical assessment, token economics, market structure, ecosystem positioning, regulatory exposure, team and governance, risk matrix, narrative analysis, supply-chain transmission. Every heading present. Every table formatted. Every cell, without exception, reading N/A. The upstream parser had failed at ingestion, returned an empty payload, and the downstream model โ to its genuine credit โ refused to invent a project. It emitted a complete skeleton and a disclaimer instead. Then something more interesting happened than the failure itself: two allocators on the distribution list read the document end to end and described it to me, in writing, as 'clean, no red flags flagged.'
That sentence is the most expensive sentence in this industry. Not because it was dishonest. Because it was structurally inevitable. A document that is 100% empty and a document that is 100% negative look identical to a reader who is scanning for red flags rather than for facts. The void does not announce itself as a void. It announces itself as silence, and silence reads as safety.
If you want to know where the next systemic failure in crypto research originates, it will not be a bad model. It will be a good model, correctly constrained, producing a correct null, and a market that has no vocabulary for a correct null.
Context: The Industrialization of Crypto Research
Between 2023 and 2026 the production function of crypto research changed more than the assets it covers. A decade ago, a token report was a human artefact: one analyst, three weeks, a spreadsheet, an argument. Today the median 'institutional-grade' research note is an assembly line โ data ingestion, entity resolution, schema templating, LLM drafting, human sign-off that averages eleven minutes. The volume is extraordinary. The marginal cost of a plausible-looking forty-table document has collapsed to roughly the price of a few thousand tokens of inference.
The regulatory layer accelerated this rather than restraining it. MiCA's operational requirements, and the parallel third-party due-diligence expectations that institutional allocators imported from traditional asset management, converted research from a discretionary input into a compliance artefact. A fund does not need a good report. It needs a report, on file, dated, with a methodology section, proving that somebody looked. That is an important distinction, and it is not a cynical one โ it is simply what compliance is: a documented process, not a guaranteed outcome.
The consequence is that the demand curve for research is now dominated by buyers optimizing for documentary completeness, not analytical accuracy. And schemas exist precisely to make completeness measurable. A pipeline that produces all forty-one tables scores 100% on completeness. A pipeline that produces one honest paragraph saying 'I cannot verify this issuer's reserves' scores 3%.
I have watched this incentive gradient up close. In 2025 I worked with legal-tech teams mapping MiCA-era arbitrage for cross-border payment firms โ seven jurisdictions with workable stablecoin treatment and credible AML supervision, a matrix comparing compliance cost against liquidity access, three fintechs that relocated to Abu Dhabi on the strength of it. That document was useful because every cell in it was fought over. Someone argued about each one. The matrix took five weeks and it was, by the standards of the format, ugly: uneven confidence levels, footnoted disagreements, one row that said 'insufficient evidence, treat as hostile.' Nobody wanted to publish that row. It was the only row that mattered.
Core: Epistemic Liquidity and the Void Ratio
Here is the framework I use, and it comes directly out of the work that got me into this business. In 2020, as a data science student, I spent six weeks building a Python mapper of Uniswap V2 liquidity depth across fifteen major pairs. The headline number โ aggregate volume โ looked like a functioning market. The depth map said otherwise: roughly 60% of observed volume was circular, self-referential, or otherwise economically meaningless. DeFi was not a liquidity revolution. It was a liquidity illustration. The order book was theatre and the tape was the script.
That experience gave me a permanent methodological bias: never trust a headline metric until you have measured the cost of moving it. Depth, not volume. Slippage, not turnover. The number that tells you the truth is always the second derivative โ how much the price moves when someone actually tries.
Apply the same logic to information instead of orders and you get a metric I have been refining for the better part of two years: epistemic liquidity โ the depth of verifiable, timestamped, attributable information available about an asset at a given moment. An asset with deep epistemic liquidity can absorb a bad headline without repricing, because the market has enough independent verified facts to contextualise it. An asset with thin epistemic liquidity moves 30% on a deleted tweet, because one marginal unit of information is a large fraction of the entire book.
The useful corollary is the void ratio: the share of a research deliverable's fields that are structurally present but semantically empty โ rendered, formatted, and containing nothing. The document I received in October had a void ratio of 1.00. Not one cell carried a fact. And because the schema was adhered to, it scored perfectly on every automated quality check we run.
That is not a bug in the pipeline. That is the pipeline doing exactly what it was specified to do, and the specification being wrong at a level nobody instrumented.
Now extend it. If you compose epistemic liquidity (how much verified information exists) with the void ratio (how much of the apparent information is empty), you get a two-by-two that explains most of the pathologies I see in 2026 research distribution:
| | Low void ratio | High void ratio | |---|---|---| | Deep epistemic liquidity | Genuine analysis; disagreements are legible | Redundant restatement; consensus echo | | Thin epistemic liquidity | Honest but fragile; a single headline reprices | The null report; silent, filed, and read as safe |
The bottom-right quadrant is where systemic risk accumulates, and it is the quadrant our entire research apparatus is optimised to produce. Thin assets are exactly the ones that get templated, because nobody has the primary data to fill the template, and high-void documents are exactly the ones that pass automated QA, because automation measures form.
The timestamp problem compounds this and it is the part that genuinely worries me. The failed report I received flagged its own gap โ no timestamp, no macro context, 'time sensitivity unevaluated.' In crypto, the half-life of information is measured in hours. If a report is twelve hours stale, its expected analytical value is close to zero. But its perceived authority decays over weeks, because it sits in a shared drive with a professional-looking cover page and nobody re-reads the date. That asymmetry โ value decaying in hours, authority decaying in weeks โ is the single most exploitable structural inefficiency in crypto information markets. It is an arbitrage, and it is being harvested continuously, mostly by accident.
There is a payments-specific version of this that I find genuinely uncomfortable. Cross-border settlement research is now heavily dependent on automated sanctions and counterparty screening pipelines. A null return in that context โ 'no adverse findings' โ is not a neutral output. It is an affirmative claim about the absence of something, generated by an instrument that may simply have failed to look. In traditional correspondent banking, a failed screen is an exception that halts the payment. In an LLM-mediated research stack, a failed screen is a blank cell that renders as a green light. I have audited enough KYC theatre to know how this ends: the honest user pays for the compliance, and the person who understands the pipeline's failure modes pays nothing at all.
The Template Is the Product
Let me be precise about the mechanism, because 'AI hallucinates' is the wrong diagnosis and it leads to the wrong fix.
A schema-driven research pipeline has two objectives that are usually aligned and occasionally catastrophically opposed: produce a document that matches the schema, and produce a document that is true. Under normal conditions, filling fields with facts satisfies both. Under data-scarcity conditions, they diverge โ and the optimisation pressure resolves in favour of the schema, because the schema is what is being measured. This is Goodhart's law with a rendering engine attached.
The constrained model in my October pipeline did the right thing: it refused. It output the void explicitly, marked it, and warned downstream. That behaviour is expensive and it is being selected against. Consider what happens at the distribution layer. A pipeline that refuses to produce output when data is missing generates exceptions, escalations, and a reputation for being 'unreliable.' A pipeline that produces a complete-looking document with a large disclaimer section generates no exceptions and a reputation for being 'robust.' The disclaimer is the tell โ I have started treating the length of a methodology-and-limitations section as an inverse indicator of content density, because a document with 4,000 words of caveats and 300 words of fact has told you exactly where it stands.
And the human layer does not catch it, because humans use form-completeness as a credibility heuristic and always have. We did this before LLMs. The pre-2023 version of the null report was the analyst who copied last quarter's model, updated the price cell, and shipped it. The difference now is throughput. A human could produce maybe two hollow documents a week. A pipeline produces four hundred a day, at consistent typographic quality, and each one enters the same shared drive with the same filename convention as the real thing.
Contrarian: Hallucination Is Loud, Completion Is Quiet
The consensus worry about AI in crypto research is fabrication โ the model inventing a partnership, inventing a TVL figure, inventing a founder's employment history. That worry is well-founded and it is also, from a risk-management standpoint, close to irrelevant. Hallucination is a loud failure. It gets caught, because invented specifics are falsifiable specifics. A fabricated TVL number can be checked against a chain in ninety seconds. A fabricated partnership generates a denial. Fabrication leaves fingerprints because it produces claims.
The real failure mode produces no claims at all, and therefore leaves no surface for falsification. A null document cannot be fact-checked, because there is nothing in it to check. It can only be judged, and judgement defaults to form. This is the blind spot in every AI-governance framework I have read applied to crypto research: they regulate outputs for accuracy, assuming outputs contain assertions. They have no provision for the output that asserts nothing and is consumed as though it asserted safety.
Here is the second-order effect, and it is the part I would want an allocator to sit with. In a sideways market โ and we are unambiguously in one, chop is the regime, positioning is the game โ the alpha does not come from directional conviction. It comes from identifying mispriced optionality before the tape resolves. Which means the marginal value of research in this regime is entirely concentrated in differentiation: what do I know that the consensus document does not?
A high-void document is, by construction, identical to the consensus document. It is the consensus document, rendered at scale. So the industry is currently spending an enormous amount of money to manufacture documents that contain precisely zero information gain, and calling the resulting process 'coverage.'
Flip it, though โ and this is where the contrarians get it wrong in the other direction. The existence of voids is not itself the opportunity. Some people will read what I have written and conclude that the play is to hunt for assets with the thinnest information coverage, on the theory that thin coverage equals mispricing. That is a coin flip dressed as an edge. Thin coverage is equally consistent with 'undiscovered' and with 'correctly ignored.' A void and a negative finding are observationally identical from the outside and completely different in expected value. Distinguishing them is the entire job, and it is not automatable, because automation is what produced the ambiguity.
This is what I was chasing in 2024 when I argued, against near-universal consensus, that the spot Bitcoin ETF would increase volatility rather than suppress it โ that active ETF traders would create a new arbitrage layer between spot and derivatives, widening basis spreads. I was told it was contrarian nonsense by people who had not back-tested 2013 to 2017. The spreads widened. The mechanism was structural, not directional: institutionalisation changes market architecture, not price. The same logic applies here. AI-mediated research does not change what is true about an asset. It changes the architecture through which truth reaches a decision-maker โ and architectures have failure modes that are invisible from inside them.
Takeaway: Instrument the Void
I do not think this gets fixed by better models. I think it gets fixed by disclosure standards, and I think those standards get written after the first enforcement action, not before.
The specific rule I would push for is unglamorous: mandatory null disclosure. Any research artefact distributed to an investment committee must carry an explicit, machine-readable statement of coverage โ what fraction of fields were populated from primary sources, what fraction from secondary sources, and what fraction were empty. Not a caveat paragraph. A number, in the header, next to the date. The date, obviously, being non-optional.
If that number existed, the October document would have read 'Coverage: 0%' at the top and nobody would have called it clean. If that number existed across the industry, the marginal cost of producing a hollow document would rise above the marginal cost of producing an honest thin one, and the incentive gradient would invert.
We already built this for liquidity. Every serious trading desk measures depth, not just volume, because a market that looks deep and isn't will destroy you on exit. Information markets have no depth metric. That is the gap, and it is a strange one, because the failure is more expensive on the information side โ you can hedge a thin order book, but you cannot hedge a decision you made because a blank cell looked like a green light.
The question worth sitting with, as we grind sideways through another quarter of positioning and no direction: when the post-mortem is written for whatever blows up next, will the proximate cause be a lie someone told โ or a document that told the truth by saying nothing, and was filed as approval?
