Over the past ninety days I ran the same test on eleven AI-assisted crypto research products. I fed each one an identical payload: a document with no ticker, no protocol name, no filing, no balance sheet, no text. A null. Valid MIME type, empty body. Ten of the eleven returned a complete, confident, fully formatted brief. One returned six words and stopped: "Input invalid. Analysis aborted."
That one is on my shortlist.
The ten failures are more instructive than any of their paid output. One report described a "modular settlement layer" with a four-year vesting schedule, a twelve-month cliff, and a governance profile rated "moderately concentrated." None of it was sourced, because there was nothing to source. The model had been handed a template and had filled it. Every sentence was grammatical. Two of them contradicted each other. Nobody had lied.
In a market that has stopped volunteering information, that is a larger problem than a bad quarter.
The research layer of crypto has been automated over the last eighteen months, and the automation landed at precisely the moment the tape went quiet. Bitcoin has spent the better part of two quarters inside a range narrow enough to make leveraged directional trades a fee-collection exercise. Perpetual funding has oscillated within a few basis points of neutral. Spot volume on the majors has drifted lower while open interest holds roughly flat โ the signature of a market holding positions rather than building them. ETF flow, which did all the narrative work through 2024 and 2025, has flattened into noise.
When price gives no signal, demand for research rises. That is arithmetic, not paradox. In a trending market a fund manager gets paid for being directionally right. In chop, the only edge is knowing something before the tape tells everyone. So the marginal dollar in this market is not chasing momentum. It is chasing discovery: which protocol lost its largest liquidity provider last month, whose unlock cliff sits eleven weeks out, which treasury is three quarters of runway away from a raise into a bidless book.
The supply side answered. Research production, historically the slowest and most labour-intensive part of the crypto stack, now runs on language models. Ingest, parse, extract entities, score across a fixed set of dimensions, publish. Marginal cost per note: effectively zero. Throughput: unlimited. And that is exactly where the failure mode lives โ not in the models, but in what the pipeline does when the input is not what it assumed.
I have watched this stack from the other side. In late 2017, before the ZRX sale, I spent a week reading the 0x settlement contracts instead of the deck, and the thing that nearly killed the position was not the code โ it was a liquidity assumption that had never been written down anywhere. I have been hostile to unwritten assumptions ever since. The current research stack is built almost entirely out of them.
Here is the mechanism, stripped of the marketing. A modern research agent is a chain of four components. An ingester pulls raw data. A parser normalizes it into a schema. An extractor maps schema fields to claims. A generator writes prose. Each component is individually reasonable. The schema has optional fields, because real data is messy. The parser returns a null for anything it cannot resolve. The generator is prompted to explain the tokenomics or summarize the governance structure.
Nothing in that chain is designed to stop.
If you have written Solidity, the contrast is immediate. require(addr != address(0)) exists because a null address is a loss event, and the EVM would rather revert the whole transaction than let it through. On-chain, an unknown value is an emergency. Off-chain, inside a language model pipeline, an unknown value is a completion opportunity. The model does what models do: it produces the most probable continuation given the surrounding text. A tokenomics section almost always contains a vesting schedule, so a vesting schedule appears.
I call this null propagation, and it is not a defect in any single component. It is the default behaviour of the system. A pipeline that cannot fail loudly cannot be trusted to succeed quietly.
The tell, once you know to look for it, is not error. It is fluency. Fabricated analysis reads better than real analysis, because real analysis is full of unresolved questions and fabricated analysis is not. One of the reports I tested carried a risk matrix with six categories, each with a probability, an impact rating, and a mitigation. It was internally coherent. It was also describing an entity that does not exist, and it never once said so.
The honest pipeline โ the one that returned six words โ did one thing differently. It placed a validation gate between parsing and generation. Empty input failed the gate. Publication halted. Nothing was written. That is a require statement in an off-chain system, and it is rarer than it should be.
We spent most of a decade hardening the price layer. Chainlink nodes with deviation thresholds and heartbeats. TWAPs computed over windows long enough to defeat single-block manipulation. Circuit breakers on venues. We built all of it because we had learned, painfully, that a number feeding an automated decision is an attack surface.
We have built nothing equivalent for the interpretation layer, and the interpretation layer now feeds automated decisions too.
The distinction matters more than it sounds. A price oracle can be wrong by a basis point and the failure is measurable โ you compare it against a second oracle and the deviation is the error. An interpretation cannot be compared that way. If a research note calls a protocol's governance "moderately concentrated" and the true top-ten holder share is 71%, no deviation threshold fires. There is no second note to compare against, because the second note was generated by the same pipeline from the same empty field.
I learned this the expensive way. When TerraUSD unwound in 2022, my fund's price feeds were not the problem. Our stableswap model assumed a curve that behaved like a constant, and the data we fed it was accurate right up to the point where the assumption was the entire position. The number was correct. The meaning was wrong. Since then I have treated the interpretation layer as a first-class attack surface and audit it with the same hostility I reserve for unaudited contracts.
Provenance is what that audit looks like in practice. For every number that reaches a memo, I want four things: the block height, the node operator, the query, and the timestamp in UTC. Not because I distrust my analysts โ because I distrust everything upstream of them. Which RPC endpoint answered? Was it an archive node or a pruned one, and did it silently serve truncated history? Which indexer, and did it handle the last reorg, or is the dashboard quietly displaying an orphaned branch? Which subgraph version, given that hosted-service sunsets have broken more DeFi dashboards than any exploit? Which exchange API, and did the v3 deprecation cut the funding-rate history in half with no error code?
Last quarter I ran a reconciliation across three independent liquidity sources for a mid-cap pair. Reported depth on the primary venue was $4.1 million within two percent of mid. The second source said $2.3 million. The third said $1.6 million. The gap was not noise. A single liquidity provider with a concentrated position accounted for the difference, and that wallet had been net-withdrawing for eleven days. Liquidity vanishes faster than hype. The venue's API is not lying โ it reports what sits in the pool at the moment of the call. It simply cannot report intent.
A three-hour indexer lag does the same thing one layer down. If a value-at-risk engine reads LP positions from a lagging indexer during a volatility window, it is measuring yesterday's risk with today's confidence, and the confidence is the dangerous part.
None of this argues against automation. It argues about which signals are worth automating. The measurements that hold up in a range-bound tape share three properties: they derive from on-chain state rather than reported state, they are cheap to reproduce, and they degrade gracefully when a source disappears.
The ones I weight highest right now: liquidity-provider concentration, measured as a Herfindahl index across pool positions rather than as headline total value locked; net contract-level flow, distinguished from exchange-reported volume; unlock cliffs, weighted by the Gini coefficient of the schedule rather than the headline percentage; treasury runway, denominated in stable assets rather than the protocol's own token; and infrastructure uptime, which is a disclosure metric rather than a decentralization metric.
That last one deserves the blunt version. Sequencer status pages on the major rollups describe the availability of a single operator. "Decentralized sequencing" has been a roadmap slide for two years, and across that period the operator set on most of these chains has not meaningfully changed. I read a sequencer status page the way I read a counterparty risk disclosure โ as a description of one entity's operating discipline, not as a property of a network.
The governance layer has the same texture. Grant committees that fund promises cannot be audited, because a promise has no artifact. Retrospective funding is different in kind: it pays for work that already shipped, which means the thing being paid for exists and can be inspected. That structural difference โ retrospective versus prospective โ is why one mechanism produces verifiable output and the other produces allocation politics. Public-goods funding works when the evidence precedes the cheque.
The marginal cost of producing a research note has collapsed toward zero. When marginal cost approaches zero, price follows, and the only surviving differentiator is provenance โ the ability to state, precisely, where a number came from.
This is the same structural problem as DeFi summer in 2020 wearing different clothes. Then, protocols advertised twenty percent yields that were mostly emissions rather than revenue, and the yield was real right up until the token inflation model met a bidless market. I rotated into stablecoin pairs and staked positions before that broke, not because I was clever about the specific tokens, but because I had done the arithmetic on where the yield came from and it did not come from anywhere. The research market has the same shape now. Coverage is not the product. Discovery is the product, and discovery that cannot be traced is indistinguishable from generation. Don't trust the yield; audit the source. Swap research note for yield pool and the discipline is identical.
Our own pipeline enforces this with three independent data paths that must reconcile before anything publishes. It sounds excessive. It is not. The reconciliation step catches roughly one material discrepancy a month, and every one of those would otherwise have entered a memo as a fact. There is also a deliberate halt: if a required field is missing, the note does not go out. In 2024, while we were building custody and reporting rails for institutional allocation work under the European framework, every attestation we produced had to be reproducible on demand for auditors. It did not take long to conclude that the internal research process deserved the same standard as the external reporting.
There is a real fix on the horizon, and it is the most interesting thing happening at the intersection of crypto and AI right now. Verifiable inference โ whether through zero-knowledge proofs of model execution, hardware attestation in trusted execution environments, or on-chain inference receipts that commit a hash of the input, the model, and the output โ turns a generated claim into a checkable one. The technology is immature, the cost is real, and most of the current deployments are proofs of concept rather than production infrastructure. But the direction is correct, because it solves the attribution problem rather than the data problem. A signed inference receipt does not make a model more accurate. It makes the model accountable for what it said, and accountability is what the research layer is missing.
Until that matures, the practical answer is boring and manual: fewer claims, each one traceable.
The prevailing response to everything above is more. More feeds, more models, more agents in an ensemble, more coverage. It is the wrong answer, for a specific reason: this is not a data problem. It is an attribution problem.
Add a second model to a pipeline that fabricates to fill gaps and you do not get accuracy. You get more throughput on unsourced claims โ and worse, the appearance of consensus when both models hallucinate from the same null. Ensemble methods work when each member carries independent error. Two language models completing an empty schema do not. They carry correlated error with different phrasing.
The other half of the contrarian case concerns what a blank field actually means in a market. An unanswered question is not neutral. It is priced as risk by whoever notices it first, and whoever notices first collects the spread. That makes coverage gaps an opportunity rather than a hole โ but only for the analyst who can distinguish "unknown" from "unknown to me." Every pipeline that papers over a null with a plausible sentence destroys precisely that information. It converts a tradeable unknown into an untradeable certainty, and it does so at zero marginal cost, at scale, in a format that looks like diligence.
There is also a version of the near future where the research layer gets cleaned up by the same force that cleaned up custody: institutional requirements. Once a mandate stipulates that research inputs carry data lineage, the number of desks able to comply will be far smaller than the number currently publishing. I would not bet against that arriving within two years.
In a tape where price tells you nothing, the scarcest asset is a number you can trace back to a block. Everything else is completion โ fluent, confident, unowned. When the next note lands in your inbox, ask the question that matters: can you name the block height, the node, and the query behind its central claim? If the answer is no, you are not reading research. You are reading a machine's most probable continuation.