The document that crossed my desk was 4,100 words long and contained zero information. Nine dimensions โ technical architecture, token economics, market structure, ecosystem position, regulatory exposure, team and governance, risk matrix, narrative, supply-chain transmission โ every field carrying the same value: N/A, insufficient information. The scorecard at the bottom awarded one star out of five across four axes. No fabricated TVL. No invented roadmap. No hallucinated audit date. Then, in the risk section, one line that justified the entire exercise: "Empty data does not mean safe. Information absence is itself a form of analytical risk." I have read maybe forty crypto research notes this month. That sentence was the only one that told me something true about the state of the industry.
The pipeline that produced it is standard in 2026. Stage one extracts information points from a source. Stage two maps those points onto a fixed nine-dimension template. Stage three distributes. The architecture is efficient, cheap, and โ this is the part nobody audits โ invariant to input. A template that always emits nine sections will always emit nine sections. Feed it a forty-page protocol specification and you get nine sections. Feed it an empty string and you get nine sections. The schema does not know the difference. Only the field values do, and the field values are strings, and strings get coerced downstream by everything that touches them.
2017 vibes. Proceed with skepticism. The ICO era taught us that the marginal cost of a token is zero and the marginal cost of a whitepaper is a weekend. We learned the lesson about tokens. We never learned it about analysis. In 2017 a bad report cost an analyst a week of their life, so bad reports were rare and expensive. In 2026 a bad report costs four cents of inference and eleven seconds, so they arrive in floods, formatted identically to the good ones, indistinguishable at the point of consumption.
Here is the mechanical failure, and it is mechanical. In every scoring system I have audited, the composite risk score is a weighted sum of flags. The weights are published. The flags are not. N/A is not a number, but a database column is typed, and the column is numeric. N/A becomes NULL. NULL becomes zero. Zero becomes "no flags raised." "No flags raised" becomes "passes screen." Four coercions, none of them logged, and a protocol nobody has examined now sits in the approved queue with the same score as one that was examined and came back clean. The queue cannot tell them apart, because by the time they reach the queue they are the same object.
I spent four months in 2022 reverse-engineering FTX's withdrawal engine. The failure was not a wrong number. It was a missing ledger entry treated as a zero balance, and a zero balance treated as a settled withdrawal. Absence was read as equilibrium. That is the same coercion, one layer down the stack, and it cost customers roughly nine billion dollars. The code was not lying. The code was answering the question it had been asked, and the question was typed wrong.
Last year I spent five months verifying the soundness proofs of a recursive SNARK implementation. The edge case was not a broken curve or a weak Fiat-Shamir transform. It was a verifier that accepted a proof of the empty statement. Vacuous truth: a universally quantified claim over an empty set is trivially satisfied. The cryptography was correct. The specification was correct. The input set was empty, and empty satisfied everything. I wrote it up for the developers directly, because the finding was not interesting to anyone who reads price charts.
Three domains, one bug: empty input is not neutral. Unless a system explicitly refuses to make a claim, absence gets silently promoted to affirmation. Accounting, cryptography, analytics โ same coercion, different stakes. The FTX version cost nine figures. The SNARK version would have cost state integrity. The analytics version costs you a position you should not have taken, and it costs it quietly, one approved queue entry at a time.
The report's risk matrix listed six categories โ technical, market, operational, regulatory, competitive, narrative โ all N/A, and then assigned a composite rating of "N/A โ insufficient information," with an explicit note that this is not the same as low risk. That distinction is the whole game. Every dashboard I have opened in the last five years collapses it. The cell is empty, the conditional formatting renders empty as green, and green means go.
The incentive gradient behind this is the one that broke liquidity mining. Subsidize a metric and the metric fills with the cheapest available substance. Subsidize TVL and you get recursive farming loops with no users at the end of them. Subsidize report volume โ forty notes a quarter, one per analyst per week โ and you get schema completion: nine sections, all present, all hollow. A desk that publishes twelve reports and writes "insufficient data" on twenty-eight is penalized for honesty. A desk that publishes forty and lets the template do the typing is rewarded for throughput. Entropy wins in research the same way it wins in liquidity pools: the system optimizes the number you are measuring, not the thing you wanted.
Layer 2 has the same shape and I have said so for three years. Dozens of rollups, the same few hundred thousand real users, liquidity sliced thinner with every launch. Research output has the same shape now: dozens of reports per day, the same four or five facts circulating, attention sliced thinner with every publish. Neither of those is scaling. Both are dilution wearing growth's clothes, and both are measured by a metric that goes up when the underlying thing goes down. Both industries share a tell, too: when the metric gets questioned, the answer is always more product, never fewer claims.
There is a one-line test for any research pipeline. Feed it an empty string. If it returns a scored report, the pipeline is a random number generator with good typography. If it returns a refusal, the schema has a null type, and that null type is a first-class citizen rather than an error state. Almost nothing in production passes. I have run this against eleven vendor products in the last eighteen months. Two passed. Both were built by people who had previously been burned by a number that turned out not to be a measurement.
Impermanent loss is real. Do your math โ and the math generalizes. A missing data point is a position. It has a value at time zero, and that value diverges from the value of the data you should have had. The divergence is monotonic in time: a protocol audited in 2023 and still marked N/A in 2026 is not the same asset as one audited last month. Stale absence compounds. Nobody marks it to market, so nobody notices the drawdown until the position is already underwater.
The only part of that 4,100-word document with real information gain was its last page โ the remediation path. It listed a minimum viable input set: title, project name, three to eight core facts, source, timestamp. Five fields. The author constrained the output to what was actually knowable and produced a specification instead of a verdict. That is correct behavior, and it is rarer than a profitable LP position. Most pipelines would have filled the five fields with plausible values and shipped.
The industry is guarding the wrong failure mode. Everyone is building defenses against hallucination โ the model that invents a TVL figure, a partnership, an audit that never happened. Hallucination is dangerous but falsifiable. You can check the number. You can call the partner. You can ask for the auditor's report. Schema completion is worse, because it manufactures a claim that cannot be checked: "no risk flags identified." Try falsifying that. You cannot disprove a missing data point. The empty report is unfalsifiable by construction, and unfalsifiable claims are precisely the ones that propagate โ into indices, into screeners, into the due-diligence packet a compliance officer signs without reading past page two. The document flagged this itself: downstream contamination. An N/A that resolves to zero does not stay local. It gets averaged, aggregated, cited, and by the third hop nobody remembers the cell was blank. And adding more data will not fix it. More input through a coercing schema produces more confident zeros. The bug lives in the type system, not the dataset.
So watch the refusal behavior, not the model size. The next cycle's real edge will not come from faster extraction or bigger context windows. It will come from pipelines that can return "I do not know" as a value rather than an exception, and from analysts who read the scorecard before the narrative. Entropy wins. Always check the fees โ and always check whether the zero in front of you is a measurement or a placeholder. One is information. The other is a liability wearing its clothes. Read the nulls first. They are the only part of the document that was written by someone who checked.