I have a document open in front of me. Nine analytical dimensions. Roughly forty table cells. A risk matrix with fifty-four entries. A scoring block that grades technical value, investment value, timeliness value and reference value.
Every substantive cell reads N/A.
That is not a flaw in the document. It is the document. And the only element in it that carries actual information is the count of times the author admitted they were looking at nothing โ fifty-one, by my tally, against a single genuine finding: the upstream extraction pipeline failed. One bit. Four thousand words of scaffolding wrapped around one bit.
I have spent twenty-seven years reading research that looks like work. This is what work looks like when it has learned to imitate itself.
Context
The bull market of 2024 through 2026 did not just reprice assets. It repriced rigor. After the 2022 cascade destroyed a generation of funds that had been underwriting on Telegram conviction, the allocators who survived demanded process. Frameworks appeared. Nine-dimension matrices. Due diligence checklists with sub-checklists attached. Every fund published its methodology. Every research desk started shipping the same six sections in the same six orders, with the same bolded headers and the same confidence annotations.
The intent was sound. The outcome is an industry that can now manufacture the appearance of analysis without the substance of it, at a scale that was impossible when analysts still had to read the code themselves.
I understand the appeal. In 2017 I spent four months inside Zilliqa's Nakamoto Consensus implementation, tracing Scilla's transaction semantics line by line. The entire published output โ twelve thousand words on my personal blog โ hung on a single edge case in finality that the team had not modeled. Twelve thousand words, one bug. The ratio matters. Had I shipped a fourteen-dimension framework instead, I would have produced more pages and less information, and the developer circles that dissected the piece would have had nothing to argue with.
In 2024 the same pattern repeated. My critique of the SEC's spot Ethereum ETF filings and the staking mechanics issuers proposed ran eight thousand words, but the load-bearing section was three paragraphs on slashing liability for custodial validators. The rest was context. Context is not padding โ but context is not the finding either. Research desks have quietly inverted that relationship. The context became the product. The finding became optional.
Core
A research report is a communications channel. It takes an input โ verifiable facts about a protocol โ and emits an output: a claim a reader can act on. The measure of a channel is not bandwidth. It is how many bits survive the trip from source to reader.
An N/A is not a low-signal message. It is a zero-bit message. Fifty-one of them arranged in a grid, under bold headers, with confidence ratings attached, still total zero bits. The formatting does not add entropy. It adds the appearance of entropy, which is worse, because it costs the reader the same attention they would have spent on real findings and returns nothing for the expenditure.
Three mechanisms generate this. They compound.
Template capture. The framework begins as a tool for organizing facts and ends as a substitute for finding them. Once a nine-dimension matrix exists, the analyst's job shifts from answering questions to filling cells. The template stops being a lens and becomes a mold. I watched this happen inside a due diligence desk in 2023: junior analysts were evaluated on coverage, not on whether any single cell contained a verifiable claim. Coverage is trivially measurable. Truth is not. So coverage won, and it has kept winning.
Null-signal inflation. When empty cells must be filled, they get filled with hedged language. "Risk level: N/A." "Probability: N/A." "Mitigation: N/A." Each entry is technically honest and collectively useless. The document swells to four thousand words because the template has four thousand slots โ not because there are four thousand things to say.
Authority laundering. This is the dangerous one. A bare "I could not verify this" is a warning. "Confidence: high" attached to the claim that inference is impossible is something else entirely. It borrows the vocabulary of rigor to dress an absence. The reader skims, sees structured tables, sees confidence ratings, sees a risk matrix with fifty-four cells, and concludes that diligence occurred. Complexity hides risk โ and a fifty-four-cell risk matrix is the purest form of complexity ever engineered to hide the fact that no risk was ever assessed at all.
I made this mistake in the opposite direction once, and it taught me the same lesson. During DeFi Summer 2020 I audited MakerDAO's V2 migration logic and found an oracle manipulation vector in the Chainlink feed integration for KNC. I published a risk assessment warning of liquidation cascades. The exploit did not fire immediately. But three major risk protocols cited the analysis, and Maker adjusted their collateral thresholds anyway. Nobody cited the framework I could have built. They cited the vector. They cited one integration and one cascade model. The insight reduced to a single question: what happens to this feed under a liquidity gap? Everything else was presentation.
The pattern is not confined to reports. It is how the industry now produces everything. In 2022 I built a six-month model of UST's death spiral โ not a framework, a differential equation, run against hourly on-chain liquidity depth. The seigniorage model had a circular dependency that could not survive sustained redemption pressure, and the curve told you roughly when. The insight was arithmetic. The credibility came from the arithmetic being checkable by anyone with the data.
Contrast that with what ships now. Prompt-to-report pipelines. Agents that ingest a topic, select a template, and emit twelve hundred words of balanced-sounding risk language with no source column. I have read a dozen of these in the last quarter. They are fluent. They are structurally sound. They contain no testable claim. And when the underlying data pipeline fails โ as it did for the document in front of me โ the report still ships, because the only thing the system was ever optimizing for was output.
That is the tell. A system designed to find truth stops when the input is missing. A system designed to produce documents keeps going. No pipeline that can complete successfully on empty input should be trusted on full input.
Contrarian
Here is where the defenders of automated research are right, and where I have to concede something uncomfortable.
An honest N/A is genuinely rare. In a bull market, the overwhelming majority of published analysis is fabricated โ not maliciously, but structurally. Numbers get estimated. Tokenomics get projected from marketing decks. Throughput figures get copied from a 2021 blog post that was itself copied from a testnet run. The fabricated claim and the empty cell look very different to a careful reader, and only one of them is safe.
The document that triggered this piece did the correct thing. It refused. It stated plainly that no conclusion could be drawn, marked every unanswerable field, and produced a list of what it would need to proceed. That refusal is the single most disciplined act in the entire crypto research stack. Most desks cannot do it. The incentive is always to fill the cell.
I learned that the hard way. My 2021 deconstruction of the BAYC contract โ centralized metadata storage, ERC-721 interoperability gaps, gas inefficiencies in the mint function โ polarized the community precisely because I published numbers instead of a thesis. Several people told me the piece was "unhelpful" because it did not tell them whether to buy. It was not supposed to. Audit the code, not the pitch. The code said what it said.
So the framework-builders have a legitimate point: structure prevents hallucination. A nine-dimension template forces an analyst to confront the dimensions they have not considered. That is real value, and it is why the format spread so fast.
The failure is not the framework. The failure is treating the completion of the framework as the deliverable.
Takeaway
The next cycle's due diligence edge will not be a better template. Templates are commodities now; every agent ships one and every fund publishes one. The edge is a source column.
Ask any research desk one question: for each claim in your last report, what is the provenance โ a transaction hash, a named contract function, a governance forum post, a regulatory filing? If the answer is a shrug, the report is a null signal wearing a suit. Trust no one, verify everything โ including, especially, the documents that most resemble work.
Sharding is easy; consensus is hard. Filling a matrix is easy. A source is hard.