Over the past seven days, a research pipeline I help oversee ingested a payload with six mandatory fields set to null. It returned eleven hundred words, a nine-dimension scorecard, four comparative peer analyses, and the word "moderate" placed next to a confidence rating. No project was named. No token existed. No funding round, no mainnet, no audit report, no publication date. The output read like every other research note in the feed that week. That is the finding, and it is not a small one.
The payload was deliberate. I blanked the fields myself, on a Tuesday, on a machine that has no network access, because I have spent nineteen years watching the same failure repeat in a different costume. In 2017 it was a whitepaper with a sixteen-page token allocation table and a two-paragraph technical section. In 2026 it is a formatted report with a nine-dimension matrix and a confidence tag. The costume changes. The structural defect does not: the form of verification is being substituted for the act of verification. And in a sideways market, where nobody can be made rich by a single directional call, form is the only thing left to sell.
The pipeline did not hallucinate a fact in the ordinary sense. It did something worse. It produced a document whose shape implied that facts had been checked. There is a name for this and I want to fix it precisely before the rest of this piece: hallucinated determinism. A wrong number can be caught. A wrong number wearing a scorecard cannot.
What the framework was for
The nine-dimension checklist is not arbitrary. I have used a version of it since 2020, when I built a Python arbitrage model that priced liquidity depth across Uniswap and Curve for a proprietary desk. The model captured roughly forty-five thousand dollars of alpha before yield compression flattened the spread, and it taught me the only lesson that has survived every cycle since: the edge was never in the model. The edge was in the input completeness. Every dimension on that list exists because a specific, painful, real-world failure traced back to a missing input rather than a wrong conclusion.
Technical dimension. You need the scheme name, the layer it sits at, testnet or mainnet status, whether the repository is open, and the audit status. Not because audits are infallible. Because the absence of an audit is a signal at least as strong as its presence, and it is the only signal most readers will ever be able to check for free. In 2017 I audited fifteen early-stage ICO contracts for the Ethereum Trust Initiative and found critical reentrancy exposure in three of them. Every one of those three had a whitepaper describing a security model. Two had named advisors with genuine credentials. None had a third-party review. The bytecode was the only witness that did not lie, and the bytecode was not in the marketing.
Token economics. Issued or not, supply, allocation, unlock schedule, and — critically — the source of the advertised yield. An APR is a number. Where the APR comes from is an argument. Those are not the same category of object and they should never appear in the same cell.
Market. Price, market cap, recent range, venue coverage, funding rate. This is the cheapest dimension to fill and therefore the one most often used to make a thin report feel dense.
Ecosystem position. Upstream and downstream integrations, developer activity, user counts. Hard to fake at scale, easy to paraphrase badly.
Regulatory posture. Domicile of the entity, domicile of the team, KYC status, and whether the token could plausibly clear a securities analysis in its primary market. This dimension is the single most common omission in retail-facing research and the most common cause of a report aging into uselessness in under a quarter.
Team and governance. Names, backers, round structure, actual on-chain governance participation. A governance forum with nine thousand registered wallets and forty-one voters is a different object than a governance forum.
Risk. Which is not a dimension at all. It is the residual of the other six, and it cannot be filled independently. Any report that assigns a risk score without six populated inputs above it has produced a random number with a decimal point.
Narrative and expectation. Which regime the asset sits in, and how far into that regime it is. Crypto narratives decay on a three-to-six-month half-life. This is the most under-appreciated source of analytical error in the entire industry, and I will return to it.
Supply chain transmission. Which sub-sector the news actually touches — miners, venues, DeFi collateral, RWA issuance, custodied vehicles. A headline about a settlement layer is not a headline about a token, and treating the two as equivalent is how contagion gets mispriced.
Nine dimensions, each requiring four to six discrete information points to be meaningful. That is a floor of thirty-six inputs for a report that deserves the name. The blank-schema payload contained zero. It still produced a scorecard in all nine rows. A framework with no inputs is not a shortened analysis. It is a different artifact entirely, and it should not carry the same label.
The six load-bearing fields
The schema has six mandatory fields. I want to walk each one, because the reason each is mandatory is the same reason its absence is invisible in the output.
| Field | What it anchors | Failure mode when null | |---|---|---| | Title | The identity of the object under analysis | Report describes a category, not an entity | | Information points | The atomic evidentiary layer | Every downstream claim becomes unfalsifiable | | Core thesis | The author's intent and incentive | Neutral-sounding text with a hidden position | | Protocols involved | Comparables, token economics, regulatory domicile | Peer analysis defaults to archetypes | | Time sensitivity | Validity window of the conclusion | Stale analysis presented as current | | Source quality | Credibility tier of the underlying claim | Order-of-magnitude error in confidence |
Start with the title. It sounds trivial. It is the load-bearing wall. Without an identified object, the report cannot be falsified, because there is nothing specific enough to be wrong about. I have watched this exact evasion happen at institutional desks, not by accident but by preference: a note on "the modular DA landscape" is unfalsifiable in a way that a note on a named rollup's blob spend is not. The categorical report survives contact with reality indefinitely because it never touches reality.
Information points come second and they are the whole game. The working rule on my desk is a minimum of five discrete, source-attributable facts before any conclusion is permitted. Five is not a magic number. It is the minimum at which a contradiction becomes visible. With two facts you can build a story. With three you can build a narrative arc. With five, you start to notice that fact four does not agree with fact two, and that disagreement is where the actual analysis lives. A report with zero information points cannot contain a contradiction, and a document that cannot contain a contradiction is not analysis. It is liturgy.
Core thesis is third, and its function is not to state the conclusion. Its function is to expose intent. A claim without an author is unauditable, because you cannot ask what the author needed to be true. When a note tells me a token is "structurally undervalued" without a named position, a named mandate, or a named horizon, I have no way to discount for the possibility that the author is describing their own book. That is not cynicism. That is the minimum standard I would apply to any sell-side note on any instrument in any market, and crypto has spent a decade arguing it should be exempt.
Protocols involved is fourth. This one is load-bearing in a way that only becomes obvious when you try to do peer comparison without it. In 2024, before the spot Bitcoin ETF approval, I published an analysis of the custodial infrastructure differences between BlackRock's IBIT and Fidelity's FBTC, focused on proof-of-reserve mechanisms and the security properties of each custody layer. That report was possible only because I had the specific entities. Strip the names and it becomes a generic essay about ETF custody that predicts nothing, and the thing it correctly predicted — settlement latency problems during the first week of trading — would have been unguessable. Specificity is not decoration. Specificity is the mechanism by which a prediction can be wrong, and therefore the only mechanism by which it can be right.
Time sensitivity is fifth, and it is where I have made the most personal errors. Crypto narratives have a three-to-six-month validity window. Not because the technology changes that fast, but because the market's expectation of the technology changes that fast. Consider the data availability layer. The thesis that dedicated DA layers would capture durable fee revenue from rollup demand was coherent, well-argued, and widely published through a specific window. The problem was never the argument. The problem was the denominator: the overwhelming majority of rollups never generated enough data volume to justify a dedicated availability layer, and the ones that did could not sustain the fee environment the models assumed. An analysis written inside that window and quoted nine months later is not a slightly outdated analysis. It is an artifact from a different market, wearing the timestamp of the current one. I have been on the wrong side of this myself. The most expensive analytical error is not a wrong conclusion. It is a right conclusion applied past its expiry.
Source quality is sixth and it is the field that changes confidence by an order of magnitude on its own. Four tiers, and they are not close together. Tier one: primary documents — protocol repositories, governance forums, regulatory filings, on-chain state. Tier two: compliant financial media with editorial liability. Tier three: crypto-native media, faster, thinner, occasionally excellent, structurally unable to be held accountable. Tier four: key-opinion-leader posts and anonymous leaks. The same information point — say, a protocol's reserve composition — carries different weight at each tier, and a report that cites tier four without labeling it tier four has committed a material misrepresentation even if every word is true. In 2017 this distinction was mostly academic. Today it is the difference between a note that survives a stress event and a note that turns out to be a paraphrase of a paraphrase of a screenshot.
The mechanism of the failure
The blank-schema output did not invent a project. It did something subtler and harder to catch. It invented the texture of verification: a comparative peer table populated with archetypes, an audit reference that named a plausible category of firm rather than a firm, a liquidity figure expressed with the confidence interval that liquidity figures usually have. In my experience the failure is systematic, and it runs in four predictable stages.
Stage one: the schema is satisfied by category. The model does not need a protocol name; it needs a protocol category. Category-level language is grammatically identical to entity-level language, which is why it passes review.
Stage two: the missing input is replaced by the distributional average of the training population. This is where the fabricated audit firm comes from. Most audited protocols are audited by a small set of recognizable firms. The model knows the shape of the sentence. It does not know whether the sentence is true, and — this is the part that matters — it has no mechanism to represent that ignorance inside a document that already has a heading for it.
Stage three: the confidence rating is computed from the fluency of the output, not the completeness of the input. This is the single most dangerous step, and it is the one no vendor I have reviewed has solved. Confidence should be a function of evidentiary coverage. In practice it is a function of how confident the prose sounds, which is a function of how much prose there is, which is a function of how long the schema is. A longer schema therefore produces higher confidence on thinner inputs. The incentive inverts the requirement.
Stage four: the report enters the information environment and gets cited as a source. Now the fabricated audit reference has a citation, and the citation has a report, and the report has a timestamp. Two cycles later, a reader in a different market reads the citation as evidence that someone checked something. This is how a truth layer degrades into a truth theater: not by lying, but by lending the format of verification to content that was never verified.
My 2022 stablecoin contagion model is the clean counterexample, and I use it because it demonstrates what input completeness buys you. After Terra, I constructed a stress test of institutional balance sheets, mapping algorithmic stablecoin redemption curves onto money market fund exposures. The model identified roughly two hundred million dollars of exposure gaps across several mid-tier hedge funds. That number was actionable because every input was a line item, not a category. It prompted a hedging directive that held through the FTX period. If I had run the same model on blank fields, it would have returned a number. It always returns a number. That is the entire problem. A model that cannot output "insufficient data" is not a model, it is a random number generator with a beautiful user interface.
What the correct output looks like
Here is the part that will read as contrarian to anyone building research tooling in this cycle, and I want to state it as plainly as the evidence allows.
The correct response to the blank payload was not a shorter report. It was not a caveated report. It was a refusal, followed by a specific inventory of what was missing and what each missing field would have changed. That refusal is a product. It has more information content than any nine-dimension scorecard, because it localizes the uncertainty precisely rather than smearing it across a document as hedge language.
I have audited enough systems to know that the ability to say "no data" is the hardest feature to ship, because it is the only feature that makes the product look empty. The commercial pressure runs the other way. An empty response does not get screenshotted. A confident nine-dimension report gets screenshotted even when six of the nine rows are archetypes. So the market selects for fluency, and fluency is orthogonal to accuracy. This is not a new dynamic in finance — it is the oldest one — but the cost of producing fluency has collapsed to approximately zero, and the cost of checking it has not moved at all.
The blind spot in the current research layer is arithmetic. If ninety-five percent of a report's content is generated from sources that are themselves paraphrases of press releases, then the marginal information gain of the report is not ninety-five percent of its length. It is close to zero, and in some cases negative, because the report's confidence tag adds a signal that was not present in the underlying material. I would rather read a four-hundred-word note with five named facts and an explicit "unknown" column than two thousand words of nine-dimension architecture. The four-hundred-word note can be falsified by Monday.
There is a second-order consequence that almost nobody prices. Information gaps are not evenly distributed, and neither is liquidity. When a market's information set is thin — when the reports circulating are archetypes rather than entity-level facts — the same thinness shows up in the order book, because market makers size positions against their confidence in the information environment. Thin information and thin liquidity are the same phenomenon observed at two layers of the stack. Liquidity dries up before the news breaks, and the reason it does is that the people providing it can no longer distinguish a verified fact from a well-formatted guess. In a consolidation tape, where price refuses to resolve, this is the only signal worth running: not where the price is going, but how much conviction the marginal provider of liquidity can actually support. Right now, across most of the mid-cap complex, that number is lower than the research volume implies.
Which brings me to the question I have been circling since 2026, when I designed a decentralized verification protocol for AI-generated content requiring on-chain attestation for data provenance. We authenticated ten thousand data points for a DePIN provider and solved, for that narrow case, the hallucination trust problem. The lesson generalizes badly in one direction and well in another. It generalizes badly because attestation proves where a datum came from, not whether it is true. It generalizes well because provenance is the pre-condition: you cannot evaluate a claim whose source you cannot identify, and today the majority of circulating crypto research cannot identify its own sources.
The real question is not whether AI can write a research report. It clearly can, and it will write thousands today. The question is who audits the analyst. In 2017 I resolved that for myself by refusing to publish market commentary without first validating the underlying contract's audit status — a code-first rule I have not broken since. That rule now needs a second clause, because the substrate has changed: no conclusion without a traceable information point, and no confidence rating without an enumerated input count. If a report cannot render its own evidence chain, its confidence tag is not a probability. It is a typographic choice.
The next cycle will not be decided by which desk has the most fluent model. It will be decided by which desk can prove, on demand and under stress, that its outputs sit downstream of actual data. Most cannot. Most will not be asked until it is expensive to find out. And the market, as it always does in the quiet middle of a range, is currently pricing the difference between those two things at zero.
