Last week a research pipeline I was asked to review returned an empty payload. Not a timeout, not a parse error, not a malformed JSON envelope. An empty array โ nine analytical dimensions, zero extracted information points, and a status field that read, in effect, "framework delivered."
The interesting part is not the failure. Failures are cheap. The interesting part is what the system was configured to do next, and what every human operator in that seat does next by default: fill the template.
Here is the specific anomaly worth tracing. The pipeline had a schema with nine required dimensions โ technical architecture, token economics, market structure, ecosystem position, regulatory exposure, team and governance, risk matrix, narrative, and supply-chain transmission. Every field was required. Every field was nullable. And nothing in the schema distinguished between a field that had been assessed and found clean, and a field that had never been assessed at all. Both rendered identically.
That is not a research problem. That is a type system problem. And in a bull market where a hundred-million-dollar raise gets a research memo inside seventy-two hours, it is the type system that decides what gets funded.
Crypto research borrowed its instruments from equity research and credit rating, and it borrowed their failure modes along with them. The nine-dimension framework is a legitimate artifact. It was built to solve a real problem: information asymmetry. In 2019 the analyst's problem was that the project knew more than the analyst did, so the analyst built a checklist to force disclosure. Coverage of a protocol meant interrogating the team, the vesting schedule, the audit reports, the validator set.
That problem no longer describes the dominant case. In 2026 the dominant case is not asymmetric information. It is absent information wearing the costume of disclosure. A project ships a documentation site, a litepaper, a fifteen-minute Spaces recording, and a Discord where the moderators answer questions with links to the documentation site. The checklist runs. Every box gets a mark. Nothing has been learned.
I have watched this happen from the inside. In 2024 I spent six weeks optimizing circom circuits for a batch ERC-20 transfer task, chasing a fifteen percent reduction in proof generation time while the Q3 launch date walked past me. That work was legible to a framework. It produced a measurable delta, a gas number, a table. The thing that actually mattered that quarter โ that the prover's witness generation was bottlenecked on a memory access pattern nobody had instrumented โ did not fit any column. It surfaced because I was reading the constraint system at two in the morning, not because a dimension asked me to.
The framework's provenance explains this. Credit rating agencies built dimension-based scoring because they needed comparability across thousands of issuers, and comparability requires fixed fields, and fixed fields require that every issuer answer the same questions. The instrument is optimized for throughput and consistency. It is not optimized for discovery. Those are different objectives, and in crypto in 2026 they have quietly become opposed.
Meanwhile the volume has gone up by an order of magnitude. Token launches, L2 mainnets, restaking layers, AI-agent identity protocols, bridge iterations โ the coverage surface expands faster than any desk can staff. So desks automate. A retrieval layer pulls the docs and the announcement. A summarization layer compresses them. A template layer distributes the summary across nine dimensions. The output is formatted, sourced-looking, and approximately empty.
The bull market sharpens this rather than softening it. In a drawdown, capital is scarce and diligence is a competitive advantage โ you can afford to pass, and passing is rewarded because the alternative is losing money visibly. In an expansion, capital is abundant and coverage is the bottleneck. A hundred-million-dollar raise generates pressure to have an opinion before the token lists, and the pressure is not to be right. It is to be present. Presence is produced by templates. Rightness is produced by retrieval. The market is currently paying for presence.
The framework did not cause this. But the framework is what makes it invisible, because a filled template is indistinguishable from a completed analysis at the level of the artifact. You have to read the evidence ledger to tell the difference. Almost nobody reads the evidence ledger.
Strip the framework down and it is a Bayesian instrument, whether or not its authors intended it that way. The nine dimensions are a prior: a structured expectation about what a legitimate protocol looks like. Evidence is the likelihood function: how much the observed facts shift belief away from that prior. The output is the posterior.
When the likelihood is empty โ when the information array is empty, when there is nothing to condition on โ the posterior equals the prior. Exactly. Not approximately. The framework returns its own assumptions, dressed in the formatting of a conclusion.
This is why "N/A โ insufficient information" is not a lazy answer. It is the only correct answer. It is the system reporting that it has learned nothing, which is precisely what happened.
The alternative, the one every desk reaches for under deadline, is to write the prior out as if it were a posterior. The protocol is technically sophisticated. The token model is well-structured with standard vesting. The team is experienced. Risk is medium. Every one of those sentences is a description of the template, not of the protocol. I have a name for this and it is not a polite one: prior laundering. The framework's assumptions get reissued as findings, with the formatting of evidence attached.
The reason it survives review is that it is unfalsifiable in the specific way that makes review useless. You cannot point at the sentence "the token model is well-structured" and say it is false, because it does not assert anything checkable. It is a mood with a citation style.
Optimizing the prover until the math screams is a discipline I learned the hard way: you push a system until it produces a number that contradicts you, and then you trust the number. Research templates are the opposite. They are optimized until they produce a document that agrees with everyone.
Trace what actually happens to a missing value inside a research system, because it is a code path, and code paths have deterministic behavior.
Stage one: the parser returns null. The field was required, so the schema coerces it to the string "N/A". Nothing is lost yet; the null is still visible as an absence.
Stage two: the summarizer encounters "N/A" and, being a language model trained on human documents, rewrites it to something more fluent. "Not assessed" becomes "no material concerns identified". This is semantic drift, and it is not malicious. It is the model doing what models do: producing text that reads like the training distribution, and the training distribution of due diligence memos does not contain many sentences that say "we do not know".
Stage three: the aggregation layer scores the dimension. "No material concerns identified" scores as low risk, because the rubric maps absence of negative findings to absence of risk. The rubric was written by someone who assumed that a thorough assessment had occurred.
Stage four: the composite rating is computed. Nine dimensions, weighted, averaged. The unassessed dimension has now contributed a favorable datapoint to a number that will appear on a slide.
Four transformations. Each one locally reasonable. The composite output is now more confident than any individual input, which is the exact inverse of how evidence is supposed to aggregate.
I have seen this bug in Solidity, and the resemblance is not decorative. An uninitialized storage pointer returns zero. Zero is a valid uint. There is no error, no revert, no log line โ the contract proceeds with a value that means "nothing was ever set" and treats it as "the value is zero". The failure is silent because the type system has no way to represent "unset" distinctly from "set to the zero value". Research schemas have the same defect. There is no type for "unassessed". There is only a string that renders like an assessment.
Tracing the gas leak in the untested edge case usually means finding a path where a variable is read before it is written. In research, the variable is confidence, and it is almost always read before it is written.
Here is the test I apply. For each dimension, name the specific artifact that would move my belief, and check whether the pipeline retrieved it. If the artifact is not in hand, the dimension is not assessed โ full stop, regardless of how much prose surrounds it.
Technical architecture. Not the docs. The bytecode diff between the deployed contract and the verified source, if there is one. The composition of the verifier set โ how many distinct entities, what the staking requirement is, whether the same operator appears under multiple addresses. The fraud proof window and what happens to in-flight messages when it expires. The commit graph over the last ninety days, because a repository with forty thousand lines and no commits since the token launch is a museum.
Token economics. The unlock cliff dates, not the percentages. The slope of the emission curve at month eighteen, because that is when incentive-funded TVL starts looking for a new home. And the ratio of real revenue to emissions-funded yield โ if the protocol's fee take is under thirty percent of what it pays out to attract the liquidity it is paying out to, the APY is a marketing number with an accounting wrapper.
Market structure. Who the marginal buyer is, and whether the float is small enough that a single unlock event reprices the asset. Ecosystem position. Whether the protocol is a dependency of something larger or a wrapper around something smaller, and how much of its TVL arrives through a single integrator. Regulatory exposure. Where the entity is registered, which is a question the docs will answer incorrectly and the corporate registry will answer correctly.
None of this is exotic. All of it is retrievable. The point is that retrieval is the work, and the framework does not do retrieval. It does indexing. It tells you which questions to ask; it cannot tell you whether the answer arrived.
I learned the distinction the hard way twice. In 2020 I spent three weeks at the assembly level of the Uniswap V2 core contracts, and the thing I found lived in an integer boundary condition the audit had classified as unreachable โ not in any summary of the constant product curve, but in the arithmetic that implemented it. And on a bridge review in 2025, the framework said: check the optimistic verification module. Fine. The finding was not in the module. The finding was in the message-passing path between Ethereum and Polygon, where a state commitment resolved before the callback that depended on it, and the ordering was not enforced by anything except an assumption in a comment. No dimension asked me to trace call ordering across two chains. I traced it because the code is a hypothesis waiting to break, and the only way to falsify it is to read the path the value takes.
The audit passed every box on the checklist. The checklist was not where the vulnerability lived.
None of this would persist if the market punished it. The market does the opposite.
A research desk is compensated for coverage. Coverage is measured in documents produced, protocols initiated, memos delivered before the token lists. Abstention โ the correct output when evidence is absent โ is invisible. You cannot put "we declined to assess this" on a slide and get credit for rigor. You get asked why the desk is underperforming.
So the equilibrium is predictable. Every desk produces a memo on every protocol, and the memos converge, because they are all generated from the same prior with the same missing likelihood. This is why so much crypto research reads identically across outlets. It is not plagiarism. It is convergent prior injection.
And the convergence is self-reinforcing, because the memo becomes an input to the next memo. Analyst A writes "technically sophisticated" from a template. Analyst B reads A's memo, cites it, and writes "widely regarded as technically sophisticated". By the third hop the claim has three citations and zero primary evidence. Each hop reduces the mutual information between the claim and the protocol toward zero while the apparent credibility increases. The system is a confidence amplifier with negative signal gain.
I ran into the arithmetic version of this in 2026, auditing a zero-knowledge credential system for AI-agent identity. The proof aggregation logic had a soundness error โ subtle, buried in how individual proofs were combined โ that would have permitted Sybil attacks on a system whose entire premise was Sybil resistance. The documentation described the aggregation as battle-tested. The documentation was correct in the sense that the code had been run many times. It had been run many times with the same flaw. Repetition is not validation, and a widely cited claim is just a claim with a distribution network.
The cycle has added a new layer to this, and it deserves naming separately because it changes the arithmetic rather than the logic.
Two years ago the failure mode was human: an analyst with a deadline and a template. Now it is a pipeline: retrieval, summarization, template distribution, and a human who reads the formatted output and signs it. The human has been moved from the position of generating the claim to the position of ratifying it, and ratification is a much weaker check than generation. Generation forces you to encounter the gaps. Ratification only asks whether the document looks like the other documents.
The summarization layer is where provenance dies. A press release becomes a summary. The summary becomes a section of a research note. The research note becomes a tweet. The tweet becomes the market's belief. At each step the text gets shorter and more confident, and the citation chain gets one link longer while the distance to the primary artifact gets one link larger. By the end, the market is trading on a claim whose nearest ancestor is a marketing document and whose furthest ancestor is a deployed contract nobody in the chain has read.
Information gain is the metric that matters here, and it is monotonically decreasing along that chain. A summary of a summary of a press release contains zero bits the press release did not contain, minus whatever the summarizer hallucinated in. The 2026 research stack is extraordinarily efficient at producing documents with no information gain and complete formatting compliance.
Latency is the tax we pay for decentralization. Formatting compliance is the tax we pay for scale. Both are worth paying when the underlying system works. Neither substitutes for the system working.
If the diagnosis is a type system problem, the fix is a type system fix, and it is boring.
First, make absence a distinct state from cleanliness at the schema level. Not "N/A" as a string in a field that also accepts "low risk". A separate status, propagated upward, that contaminates the composite. If three of nine dimensions are unassessed, the composite rating must be not a number. Not a lower number. Not a number.
Second, require an evidence pointer per claim. Every assertion in the output cites an information point index. A claim with no pointer does not render. This is mechanical, it is enforceable, and it kills prior laundering at the point of writing, because you cannot type the sentence "technically sophisticated" without a retrieval that supports it.
Third, fail closed on evidence volume. Below a threshold โ five points, ten, whatever the desk sets โ the pipeline refuses to emit an assessment and emits a request for material instead. The request is the product. It is more useful than the memo, and it is honest about what the desk does not know.
I keep an evidence ledger for my own work. Every claim I publish traces to an artifact: a contract address, a commit hash, a registry filing, a line in a specification. When the ledger is thin I write about the thinness. That is not modesty. It is the only way the analysis stays falsifiable, and falsifiable analysis is the only kind worth paying for.
Modularity isn't free; it's an entropy constraint โ you can decompose a system into clean parts, but the interfaces between the parts carry the risk the decomposition was supposed to eliminate. Research frameworks have the same property. Nine clean dimensions, and every real finding lives in the seam between two of them.
Here is the counter-intuitive part, and it is the part the industry gets wrong in the direction of optimism.
The consensus reading of an empty payload is that it is a neutral state โ nothing known, nothing concluded, no harm done. That reading is wrong. Absence of information is an absorbing state, not a neutral one. Once a dimension has been marked unassessed, every subsequent process treats it as assessed-clean, and the error compounds through every downstream artifact. The null does not sit quietly. It decays into a favorable value within two processing steps, and the decay is irreversible because nobody reopens a completed memo to check whether the evidence behind it ever existed.
The second blind spot is precision theater. Nine dimensions, each scored, weighted, composited into a number with one decimal place. That number looks like a measurement. It is not. It is a weighted average of nine judgments of wildly different reliability, three of which were generated from a template and one of which was generated from a documentation site. The decimal place is the tell. Real measurements know their error bars. Template composites do not have error bars, which is exactly how you can identify them.
The third blind spot is that the framework was built to catch bad projects and nobody built the instrument to catch bad analysis. The meta-risk is unowned. A due diligence framework interrogates the protocol's trust assumptions and never interrogates its own. Which means the most dangerous document in a 2026 data room is not the unaudited contract โ everyone knows to look at that. It is the research note that scored the unaudited contract as low risk because the audit dimension returned empty and empty scored as clean.
The edge case here is not in the protocol. It is in the analyst's tooling, and it has been shipping for years.
The forecast is not complicated. Research provenance becomes a first-class artifact. Signed evidence chains, claims bound to retrievals, composites that refuse to resolve when the likelihood is empty โ not because regulators demand it, but because the desks that adopt it will be the only ones whose memos survive a post-mortem. Coverage that cannot cite is coverage that should not ship.
The uncomfortable question is one nobody in the seat wants to answer. When the ninth dimension comes back empty and the memo is due in four hours, the framework will not stop you. It was never designed to. It will hand you a template with nine blank fields and the formatting of a conclusion, and the sentence you write into that space is a decision, not an observation.
Debugging the future one opcode at a time is slower than filling a template. It is also the only version of this work that has ever been worth anything.