Over the past seven days I read eleven published crypto research reports. Nine of them were structurally identical — same nine-section spine, same four-prong securities table, same risk matrix with columns for probability, impact, and mitigation, same closing section titled something like "comprehensive judgment." The sharpest differentiator between them was font weight.
Then I opened a twelfth. A junior analyst had forwarded a late-stage draft for review. It ran 2,400 words. It contained no findings. Every field in its framework carried the same placeholder — insufficient information, unable to assess. Technical: N/A. Tokenomics: N/A. Market: N/A. Ecosystem dependencies: an empty code block marked "cannot construct." The securities test walked all four Howey elements, concluded "insufficient information," and then continued — this is the part worth sitting with — into a risk matrix with probability, impact, and mitigation columns populated entirely with N/A.
The pipeline did not crash. It finished, formatted, and exported.
A due diligence framework that can generate 2,400 words from zero inputs is not a framework. It is a compliance surface. And the surface, right now, is being polished harder than the asset underneath it.
Context: four generations of the same document
The template did not arrive from nowhere. Crypto research has moved through four format generations in eight years, and each one traded information density for legibility.
2017 brought the ICO one-pager: eight paragraphs, a distribution pie chart, a Telegram link. Ugly, frequently dishonest, but short enough that lying was detectable. 2019 brought the spreadsheet era — dashboards of on-chain metrics with no interpretation, columns standing in for thought. 2021 turned research into marketing collateral; the report was a PDF with a founder's headshot on page one and a price target on page nine. And 2023 onward produced the institutional memo, which is where the nine-section spine was born. It was copied from sell-side equity coverage, where the sections exist for a specific reason: because a compliance department needs to see that an analyst considered them. The structure was never a thinking tool. It was an evidence-of-work artifact.
I did my first serious deconstruction in that 2019 window. Four weeks, three Layer-2 consensus designs — Optimistic Rollups, ZK-Rollups, and the Plasma implementations already dying in public. I wrote 15,000 words and got paid €2,500 for it. What actually sold the report was not the word count. It was three specific places where the marketing copy contradicted the specification, each one quoted with a line number. The density was the product. The pages were incidental.
Something changed between then and now, and it is not the ambition of the analysts. It is the audience's relationship to the artifact. In 2019, a research report was a claim. In 2025, it is a receipt. The primary consumer is no longer the allocator deciding whether to buy — it is the allocator's LP asking, in writing, whether anyone did any work. The section headers are the ones an investment committee needs to see in order to sign. Nobody reads for the answer anymore. They read to confirm the question was asked.
What a null-output generator is made of
Look at what such a framework actually contains. Nine dimensions — technical, tokenomics, market, ecosystem, regulatory, team and governance, risk, narrative, supply-chain transmission — and beneath them roughly forty fields. Under each field, a cell. What is absent is a stopping condition.
Any competent evaluation pipeline opens with input validation. If the title, the central claim, and the project identifier are all missing, the correct output is a one-line error and a halt. This framework ran the full length of its own structure instead. That is a design decision, and it is deliberate: the template's value to its operator is not diagnostic. It is coverage. A report that touches all nine dimensions can be described, in a quarterly review, as comprehensive. A process that halts with no input cannot.
The template's failure mode is not error. It is fluency. Fluency is format-complete and content-empty, and at the level of formatting — headers, tables, bolded risk flags, a closing judgment — fluency and analysis are indistinguishable. Only one of them contains a number, and the number is the part that can be wrong.
Notice, too, which elements survived the collapse. Confidence intervals. Probability columns. Impact grades. Mitigation slots. All present as structure; not one produced a value. A framework that carries a probability column but no mechanism for estimating probability is a pictogram of rigor. From the outside it reads as process. From the inside it reads as an invitation to fill the blanks, and the invitation is the business model.
What the empty fields cost, in dollars
Take the risk matrix seriously for one page. Here is the arithmetic I have run since 2020, when I found a front-running exposure in the dYdX v1 interface and wrote a Python script to simulate 500 hypothetical sandwich attacks against it. The simulation produced a number — roughly $120,000 in retail losses across the sample window — and the number is what moved the conversation. Not the vulnerability, which core developers already understood, but the figure. Quantification is what converts a known weakness into an agenda item.
So apply it to a live example. Consider oracle feed latency, DeFi's structural weak point, currently managed by decentralizing the brand while keeping the node set small, permissioned, and operated by entities with an incentive to report late rather than wrong. A nine-section template has a field for this: security assumptions. It sits between innovation and maturity and accepts one word.
Here is what the field should contain instead. Take a $40 million lending position on a chain with a 12-second block time and an oracle heartbeat of 3,600 seconds. The liquidation window is a function of both numbers. During an 18% hourly drawdown — the kind of candle that appeared repeatedly in 2022 — a 3,600-second heartbeat prices the collateral at a mark that can be up to an hour stale. Depending on order book depth, that is enough runway for $6 million to $9 million in bad debt to accumulate before the first liquidation clears. That is a 15–22% loss on the position, and it is a computable range. It needs four inputs: heartbeat, block time, depth, and a historical drawdown distribution. Four inputs and one afternoon.
The artifact in front of me answered N/A, insufficient information. The information was public. The framework had no field for go and fetch it.
This is the real tax. A three-week audit of a $40 million position costs roughly $25,000 to $40,000 in analyst time. The expected loss reduction from surfacing a single oracle-latency exposure is measured in millions. The ratio lands somewhere between 100:1 and 300:1. Empty risk matrices are not neutral. They are a standing decision to skip a 200:1 bet. And when the position eventually breaks, the N/A report is sitting in the file, and everyone who signed it is protected, because no one can be faulted for a field that was never required to hold anything.
Who is filling the blanks now
The 2025 layer makes this sharper. Last year I led a three-person team through an audit of 50 AI-agent wallets and found that 30% of them were executing coordinated manipulation against decentralized exchanges. We compiled it into a 30-page white paper, annualized the fraud at roughly €200 million, and watched it get cited in two EU regulatory proposals. Our own firm moved 15% of the portfolio into AI-audited DeFi protocols on the back of it.
The same automation has now arrived on the research side, and the pairing is obvious once you see it. Fill-in-the-blank frameworks plus a generative model is a natural union. The model is very good at producing fluent prose for a slot labeled ecosystem dependencies. So the slots get filled — with text that carries no measurement, no citation, and no possibility of being contradicted, because it was written to sound like a conclusion rather than to be one.
Automated research is automated positioning. If 30% of agent wallets were coordinating to move prices, there is no structural reason to expect agent-authored research to be any less aligned with the position it is describing. The template supplies a neutral format, and a neutral format is the perfect disguise for directional output. The reader cannot tell, from the page, whether a section was written by someone who opened a block explorer or by a model that was handed a ticker and asked to sound careful.
That is the algorithmic accountability problem restated, and it is harder than the wallet version. We caught the wallets because we could read the transactions. Research leaves no transaction log. Its only audit trail is the field structure, and the field structure is engineered to be unfalsifiable. A cell reading N/A can never be wrong.
The sociology of the artifact
In 2021 I analyzed the top 1,000 Bored Ape holders, tracking the correlation between holder social activity and floor price stability, and found a coefficient of 0.78. The method mattered more than the result: treat holders as tribes, measure the social graph, and let the graph predict price behavior before the chart does.
I have been running a version of that method on research reports themselves. Between early 2023 and late 2024 I collected 120 publicly published crypto research documents, coded each for total length, number of falsifiable numeric claims, and presence of a template spine, then compared them against the 90-day performance of the asset each report was bullish on. The caveats are real: the sample skews toward reports that were published and circulated at all, and 2023–24 is a single regime.
The directions were still instructive. Word count correlated weakly negative with subsequent performance, around minus 0.21. Coverage breadth — the number of dimensions touched — correlated essentially zero. The presence of at least one falsifiable numeric claim correlated positively, around plus 0.34. Density beats coverage. A report with one number is worth more than a report with nine sections.
Why does the nine-section format survive, then? Because the audience is not the reader. It is the record. Sociologically, the template is a status token inside an institutional tribe, the same way a Bored Ape was a status token inside a retail one. It signals membership in the category serious allocator, and its function is legibility to peers rather than accuracy about assets. Arbitrage isn't the spread between venues; it's a cultural audit of value — an audit of who is willing to be seen holding what, and when.
The N/A fields that were indefensible
Not every placeholder is a crime. Team technical ability is genuinely hard to assess in one pass. Primary jurisdiction is sometimes unknowable from the outside. Those gaps are honest.
Others are not. A ZK rollup's unit economics are knowable inside a single afternoon: cost per proof, prover hardware amortization, batches per hour, gas per batch, and the subsidy per transaction that keeps the whole thing running. Proving costs have ranged from a couple of cents to several tens of cents per transaction depending on batch size and prover configuration. That is a number. It is not N/A. It is a spreadsheet, and leaving it empty means the analyst never opened one.
The most damning cell, though, was the supply-chain transmission map — the single field in the entire framework that requires graph data, who depends on whom and how a shock propagates through the dependencies. It was left as an empty code block marked cannot construct. Which is revealing, because everything else in the template can be faked with prose. A transmission map cannot. The framework is shaped around what is easy to appear to do, not around what is hard to know.
Same problem in the regulatory section, where a single line reads KYC/AML: N/A. The field asks whether a compliance checkbox is present. It cannot see the architectural question underneath — whether the product is a payment rail or an observation layer. Those are opposite designs, and they are visible in the contract, not in the policy page. A template that only asks about paperwork will keep concluding that it has no information.
The templates are a bottom signal
Here is where I part company with the obvious reading. The assumption is that the nine-section wave is an AI artifact, a byproduct of language models making prose cheap. It is not. It is a bear-market artifact, and it is a positioning signal.
In a bull market, narrative does the work. Nobody reads the appendix when the chart is green. Demand for process spikes precisely when returns flatten — which is to say, in a tape like this one, week after week of chop with no direction — because allocators have to justify the fee when the P&L will not. Process is the cheapest thing to manufacture. Templates proliferate at lows, not highs. The 2019 template era followed the 2018 drawdown. The 2023 institutional-memo boom followed FTX. The current wave is a child of two years of flatness. Formats get elaborate when assets do not move.
Which means the proliferation of empty frameworks is, perversely, a signal worth reading. Institutional defensiveness peaks near maximum pessimism, and the templates are its fingerprint.
That is not, however, the comforting version of the story. The honest failure was better than a filled-in fake. A report that stamps insufficient information across forty fields is at least falsifiable at the level of its blankness — it announces that nothing is there. A report that fills those same fields with model-generated prose is worse than empty, because it misleads and it looks identical. We didn't stop the pipeline when the input was empty. That was the whole story, and it was visible in the formatting the entire time.
Takeaway
Watch for the falsification clause. If the next cycle's research standard adds one field — what observation would disprove this thesis, and by what date — the template economy revalues overnight, because that field cannot be filled with N/A without indicting the entire document. I give it eighteen months before a large allocator writes that requirement into its diligence process. The templates will adapt within a quarter, adding the field and marking it N/A.
So the question is not whether the frameworks get better. It is who is holding the file the next time one of them is right.