Last month a fund manager forwarded me a due-diligence report. Forty-one pages. Nine analytical dimensions. Twenty-three charts. Zero backreferences to primary data.
I read it three times. Then I tried to falsify a single claim and could not. Not because the claim was robust — because it was constructed so that no chain explorer, no commit hash, and no unlock table could ever contradict it.
He paid eighteen thousand dollars for it.
I have spent eleven years watching this industry, the last six as an audit partner. I have signed off on contracts that moved nine figures, and I have killed contracts that moved none. The document above is not an outlier. It is the median.
Crypto research industrialized between 2021 and 2023. Subscription desks replaced Twitter threads. Fund LPs demanded comparability, so desks adopted a nine-dimension scaffold lifted wholesale from equity research: technical, tokenomics, market structure, ecosystem position, regulatory exposure, team governance, risk, narrative, and supply-chain transmission.
The scaffold is sound. Each dimension, executed properly, terminates in an artifact — a block height, a merged pull request, a court filing, a vesting contract, a function signature.
What changed is where the scaffold sits. It stopped being a checklist you run against evidence and became the evidence itself. A desk that cannot fill a dimension writes "insufficient data," counts the paragraph, and moves on. Nine dimensions, nine headings, four thousand words, zero information gain.
That arithmetic is worth examining.
Every claim needs a pointer. A claim without one is not a weak claim — it is a fabricated claim wearing the grammar of a supported one.
I keep a personal standard for this. A sentence in a research note either resolves to a numbered reference — an address, a transaction hash, a timestamped filing, a line number in a repository — or it does not belong in the document. I mark the failures. I do not soften them.
Run that filter across the industry and the output collapses. I tested it. I pulled twelve reports from four subscription desks, the ones with waitlists, and deleted every sentence containing neither a number nor an on-chain reference. The survivors averaged twelve percent. Twelve. The remaining eighty-eight percent was connective tissue: adjectives, framing, and restatements of the token's own marketing copy.
Trust is a variable; proof is a constant.
There is a second failure, and it is structural — unfalsifiability by construction. "Strong fundamentals." "Experienced team." "Growing ecosystem." Each of those sentences survives every possible world. A token can lose ninety percent of its liquidity and the phrase still reads true to whoever wrote it. Ask for the falsifier. If you cannot name the observation that would kill the sentence, the sentence carries no information. Delete it.
Confidence laundering is the subtler variant. I keep seeing tags like [Confidence: High] attached to claims with no data source beneath them. Confidence is derived, not asserted. In my 2020 review of Curve's early stablecoin pool math libraries, the confidence came from four weeks of reading integer arithmetic — from proving that three specific overflow vectors existed in the documentation before launch. I submitted those through the private bounty channel. No announcement, no thread. The evidence was the bytecode. That is what a confidence rating is supposed to be shorthand for.
The pattern that produces the forty-one-page reports is this one: the template treats volume as a deliverable. Nine sections marked "UNVERIFIABLE — requires on-chain data, audit report, or unlock schedule" is not conservative analysis. It is four thousand words of noise with a compliance aesthetic. The honest version is shorter. It reads: I cannot evaluate this, here is what I would need, and here is why the absence itself is a signal.
I learned that in 2022. While the market argued about narratives, I spent seventy-two hours tracing Anchor Protocol's yield distribution — inflows against outflows, block by block. The conclusion was not complicated. The yield was debt, not revenue. The report ran forty pages because the balance sheet required forty pages, not because the format did. Regulators cited it months later, and what they cited was the flow table, not the prose.
Same method, late 2022. FTX. Four and a half billion dollars of user assets across five chains. I traced them by hand and clustered the wallets. Fourteen distinct clusters linked back to personal accounts. Every one of those clusters is a falsifiable object. Any defense counsel could check them. That is the point.
2023, the Azuki spin-off ecosystem. Sixty percent of reported volume traced to a single entity operating fifteen wallets. I did not write about the art. I wrote about the funding graph. The correlation between thin liquidity and volume spikes was the finding, and it was checkable by anyone with an explorer.
Earlier this year, an AI-agent autonomous wallet protocol. A logical race condition in the reinforcement-learning reward function. Under a specific market state, the function permitted unbounded minting. I reproduced it on testnet before mainnet. The claim is falsifiable by anyone who runs the state. The model's intentions are irrelevant. Only the determinism of the contract matters.
Five cases. Five claims. Five pointers. None of them required an opinion about the future.
Now consider what a generative model produces when you hand it the same nine-dimension template. It fills it. It produces plausible structure — headings, subheadings, hedged language, a risk table. Structure is legible. Structure looks like diligence in a PDF. But the sampler generating the prose has no access to the chain, so the pointers, where they appear at all, are plausible rather than true. That is the emergent risk of the last two years: not hallucinated facts, which are easy to catch, but hallucinated citations, which are not.
A report without backreferences is a fiction with footnotes.
Here is what the template builders got right, and I will defend it.
Before the scaffold, analysts forgot things. A 2019-era tokenomics section routinely omitted the vesting cliff entirely, because the cliff was boring and the narrative was not. The nine dimensions exist to prevent omission. An LP comparing twelve deals needs the same twelve fields. Comparability has real value, and the desks that standardized it did the market a service.
The refusal posture also deserves credit. When the information set is empty, the correct output is a refusal. A machine that generates nine fields of "insufficient data" anyway is not assisting you — it is supplying the appearance of coverage. Silence is a publishable output. Naming what you cannot verify is a finding, not a gap in the work.
So the failure is not the framework. The failure is the inversion. The scaffold moved from the back of the document to the front of the product. Diligence became a layout decision.
There is a quieter cost. Template thinking trains readers to expect completeness. When a desk publishes a genuinely narrow finding — one claim, fully evidenced — it looks thinner than the forty-one-page alternative. Information density and perceived rigor are inversely correlated right now. That is a market failure, not an epistemic one.
We are in a sideways tape, and sideways tapes are where positioning is built. The readers who need signals are not asking for more sections. They are asking which numbers hold up.
Three questions worth sending to any desk before renewal. Show me the pointer for your three strongest claims. Name the single observation that would falsify your central thesis. Count the sentences that survive when you strip everything unpointed, and tell me the percentage.
Most desks will answer none of them. The ones that answer all three are the ones worth paying.
Determinism is a feature, not a constraint. Ask for it.
