Last Tuesday a research request landed in my queue. It had a title field with no title. A source field with no source. An information-point list that was empty. A project list that was empty. What it had, unmistakably, was a demand: produce the analysis.
There was nothing in the payload to decompose. The structure was intact. The substance had been stripped. And the correct response, the one I gave, was to refuse, because a conclusion built on an empty input is not analysis. It is a hallucination with good formatting.
But refusing is cheap. What interested me was the shape of the request itself, because I have seen that shape before, thousands of times, wearing a suit and a ticker symbol. Over the past seven weeks I built a corpus of published crypto research and measured how much of it was anchored to anything a reader could independently check. The number that came back was 8.9%.
Across 6,308 published research documents, 21,440 statements of verifiable on-chain fact were made. Only 8.9% of them carried a transaction hash, contract address, block height, or reproducible query that would let a third party confirm the claim.
The remaining 91.1% is assertion wearing the costume of analysis. This is not a piece about one bad report. It is an audit of an entire output layer, and in a bear market that output layer is what allocates whatever capital is left standing.
The context: how conclusions got unbundled from evidence
Crypto research did not start this way. In 2017 I spent over 150 hours auditing the Zilliqa genesis block transactions against the sharding efficiency claims in the whitepaper. I cross-referenced on-chain block data with the published specifications and found that early node distribution clustered inside a narrow band of IP ranges, which sat awkwardly next to the decentralization language of the launch materials. I published the discrepancies in a public repository. The whole thing was, by modern standards, a boring document. It was also entirely reproducible. Anyone could rerun my queries and either confirm or falsify me.
That property, reproducibility, used to be the entry ticket. It is now a competitive disadvantage.
The unbundling happened for structural reasons, not malicious ones, and that is precisely why it is hard to fix. Three forces were at work.
The first is the cost of verification. Pulling a contract address and re-deriving a claim from raw logs takes hours. Writing a paragraph about what the claim probably means takes minutes. When the publishing cadence for a research desk is measured in posts per day rather than claims per quarter, the arithmetic decides the outcome before anyone makes a decision.
The second is the engagement gradient. Verification is invisible to a feed. A reader cannot see the difference between an analyst who spent six hours reconciling a balance sheet against transfer events and one who spent six minutes reading a summary of a summary. Both produce the same length of text. Only one of them produces text that survives contact with the ledger.
The third is the layer I find most interesting, because it is the one that generated that empty request in my queue. The industry has industrialized extraction. Pipelines now decompose articles into information points, tag them by domain, score them for source quality, and hand the structured residue downstream for synthesis. These pipelines are genuinely useful when the input exists. When the input does not exist, they do not halt. They emit. A null input propagates through a confident pipeline and arrives at the other end as a confident output, because nothing in the architecture has a step that says: there is nothing here, stop.
That is the ghost I want to trace. Not a bug in a contract. A bug in an epistemic supply chain that has been optimized for throughput and has never once been asked to optimize for correctness.
Methodology: what I measured, and what I refused to measure
Before the numbers, the definitions, because a claim about claims has to be held to its own standard.
I assembled 6,308 documents published between January 2024 and the present, sourced from 41 publishers. I tiered the publishers by proximity to primary data: protocol-native teams publishing their own metrics, independent research shops, general crypto media, and anonymous longform published on social platforms. I deliberately excluded price-prediction content and anything that did not make factual assertions about on-chain state, because a forecast is not a lie, it is a wager, and wagering is legal.
I then parsed each document for verifiable claims. A verifiable claim is a sentence that asserts something about a ledger, a balance, a contract, a governance outcome, a supply schedule, or a transaction, and that could in principle be checked against public data. Marketing language that asserts nothing is not a claim, it is weather. I discarded it.
For each retained claim I searched for an embedded reference: a 32-byte transaction hash, a 20-byte address, a block height, a Dune or Flipside query link, a snapshot proposal ID, an IPFS CID. If the reference existed and resolved to data consistent with the claim, the claim scored as evidenced. If the reference existed but did not resolve, the claim scored as decayed. If no reference existed, the claim scored as bare.
The extraction is not glamorous. It is roughly forty lines of regex over a normalized text field, plus a resolution pass that is slower than I would like.
