Nine Dimensions of N/A: The Empty Report That Exposed Crypto Research's Hallucination Problem

The Report With No Facts In It

On a Tuesday morning in early March, a founder I have mentored since her first grant application forwarded me a document and asked one question: should she size into it?

The document ran eleven pages. It carried a cover page, an executive summary, nine analytical dimensions, twenty-three tables, a risk matrix color-coded from green to red, and a confidence declaration at the end. It looked, at a glance, exactly like the research that used to cost a mid-five-figure retainer and three weeks of an analyst's life.

It contained no facts. Not one. Every cell in every table read N/A. The technical section declined to determine whether the subject was a layer-one, a layer-two, or an application. The tokenomics section declined to state a supply. The risk matrix listed six categories and, beneath each, the phrase "unable to identify." And yet the document was structurally flawless. It had a table of contents. It had a methodology note. It had a disclaimer, a set of professional glossary annotations, and a closing call to action addressed to whoever had submitted the input, politely asking them to re-run phase one with valid data.

She read it as a verdict. She believed N/A meant "no risk identified."

That moment — a well-formatted document being mistaken for knowledge — is the most consequential thing happening in crypto research right now, and almost nobody is pricing it. Over the past eleven months I have reviewed 214 machine-generated token briefs for the education platform I run in Cape Town and for two grant committees that asked for my eyes. Sixty-one of them contained at least one table where every substantive cell was empty. Nine were, like this one, entirely empty and entirely confident in their presentation. In a sideways market where everyone is waiting for direction, that is not a formatting problem. That is an information-integrity problem wearing a research report's clothes.

How We Built an Industry That Rewards the Appearance of Knowing

The instinct will be to file this under "AI slop" and move on. That filing is wrong, and the reason it is wrong matters.

Slop is easy to catch. It has the tell-tale cadence — the breathless adjective stacking, the invented partnership, the number with no unit. This was different. This document was honest at the level of every individual cell and dishonest at the level of the whole. No single statement in it was false. The aggregate was a lie of omission so complete that a reader under time pressure could not distinguish it from a thorough analysis that happened to find nothing alarming.

To understand how we got here, you have to understand what changed in the demand side of crypto research, not just the supply side.

Between 2017 and 2021, research in this industry was a literacy project. When I served as lead community liaison for MakerDAO's early Cape Town cohort during the ICO mania, I watched 500-plus speculative tokens launch in a single cycle, and my job was not to rate them — it was to explain to non-technical investors why an unbacked stablecoin is a promise with no collateral behind it. I organized twelve town-hall webinars, manually vetted 200-plus community submissions, and learned something that has never stopped being true: the marginal crypto participant does not need a price target, they need a risk map. Financial literacy is a human right, not a privilege, and research that does not transmit literacy is entertainment with a spreadsheet attached.

Then the exchange-traded funds changed the shape of demand. Post-approval, Bitcoin's marginal holder is no longer someone who read the whitepaper and internalized peer-to-peer electronic cash. It is an allocation committee that needs a one-page memo justifying a 1% sleeve. The memo does not need to be true in any deep sense; it needs to be defensible in a meeting. Templates were built for that meeting. Templates reward completeness. Completeness is cheap to fake, and the moment completeness became the product, the industry began optimizing for the appearance of knowing rather than the state of it.

Layer-two research illustrates the drift perfectly. For two years, the phrase "decentralized sequencing" has appeared in more decks than it has in more code. When a brief lists the sequencer field as N/A, the honest reading is not "unknown." The honest reading is "us." A single operator ordering transactions, with an upgrade key held by a three-of-five multisig and no disclosed timelock, is not a decentralization question awaiting data. It is a decentralization answer that the template declined to write down, because writing it down would have made the report unhelpful to the person who commissioned it.

Code is law, but ethics is conscience. A template that produces neither is just furniture.

Anatomy of a Hallucinated Report

What I want to do here is not complain. I want to take the failure apart the way I would take apart an unvalidated contract, because the failure is structural, it is reproducible, and it is now propagating through systems that will act on it without a human in the loop.

The Pre-Flight Gate

Strip away the formatting and the document that landed in my inbox was a pipeline with a missing null check.

In Solidity, we internalized this lesson a decade ago. A function that accepts an address without a require(input != address(0)) will eventually write to the zero address, and the tokens are gone. A lending market that consumes an oracle price without validating freshness and non-zero will eventually read a stale zero and liquidate an entire book of otherwise healthy positions. Nobody calls that a market event. They call it a bug, and they write a guard against it.

Research pipelines have the same obligation and almost none of the discipline. The correct behavior when the input fact-list has length zero is not to emit a nine-dimension framework populated entirely with N/A. The correct behavior is to halt, return a typed error, and refuse to produce a deliverable. That is what fail-fast means. What actually happened is that the pipeline ran to completion, because every stage of it was designed to produce output, and no stage was designed to produce a refusal.

The most dangerous artifact in crypto research today is a document that has been rendered successfully from no input. It carries every visual signal of rigor — tables, dimensions, glossary, disclaimer — while carrying none of the substance. And because it is visually indistinguishable from a low-confidence analysis at a glance, it inherits the credibility of the format rather than earning the credibility of the content.

I have started recommending a single gate to every team that asks me to review their research tooling: if the fact list is empty, or the core thesis field is blank, or the source attribution column has no entries, the pipeline must exit non-zero. Not warn. Not annotate. Exit. A pipeline that cannot say no cannot be trusted to say yes.

N/A Is Not Neutral

Here is the part that cost my founder three days of sleep.

N/A reads as neutral. It sits in a table cell where a risk rating would go, and the human eye — trained by thousands of forms — interprets an empty field as "nothing to report." That is the opposite of what it means. In an analytical context, N/A means the subject declined to disclose, or the analyst could not verify, and the correct epistemic state is indeterminate. Indeterminate is not a midpoint between good and bad. It is a separate category, and in a regulated context it carries its own weight.

Take the Howey analysis, the four-element test the U.S. Supreme Court established for whether an arrangement constitutes an investment contract. Money invested, in a common enterprise, with an expectation of profit, derived from the efforts of others. When a token brief lists all four elements as N/A, a reader can walk away thinking the security question is unresolved. It is not unresolved. It is unanswered, which in enforcement terms is worse. An undisclosed allocation schedule is not a neutral fact about a token; it is a fact about the issuer's willingness to disclose. The absence of the number is itself the number.

I have watched this play out at the governance layer for nine years. The pattern is consistent: a project publishes a governance forum, a snapshot page, and a proposal template, and the brief records the governance dimension as "community-driven." Then you open the treasury contract and find a three-of-five multisig with the same three signers who appear on the foundation's incorporation documents, and a timelock that exists in the documentation but not in the bytecode. The DAO is not a lie. The DAO is a compliance shield, and the template that recorded it as "community-driven" did not lie either — it simply declined to open the contract.

The Provenance Gap: 10,000 Versus 1,200

The single most valuable discipline I brought from my audit work into research review is a refusal to accept a number without its measurement context.

Consider a performance claim in a typical layer-two brief: throughput of 10,000 transactions per second. That figure is almost always the project's own, derived from a synthetic benchmark on a controlled sequencer with a single client, no state contention, and no calldata compression cost. It is not a lie. It is a measurement of a system that does not exist in production. Third-party observation of the same network under organic load, with real state contention and a live mempool, typically lands an order of magnitude lower — and the honest analyst writes down both numbers and explains the delta.

The template writes down neither. It writes N/A, because the fact list it was given contained only the project's self-reported figure and no third-party corroboration, and the pipeline had no rule for what to do with an uncorroborated claim. So the claim vanished into a null, and the reader received a report that was simultaneously more honest and far less useful than the marketing deck it was supposed to critique.

Provenance is the whole ballgame. A self-reported metric with no cross-validation is not data; it is positioning. An explorer-verified balance, a decoded contract call, a signed block with a timestamp — these are data. My rule of thumb after eleven months of this work: if the fact list does not distinguish between what the project said and what a third party measured, the report's confidence level should be downgraded across every dimension at once, not just the one where the discrepancy appeared. One uncorroborated claim is a signal about the sourcing discipline of the entire document.

The Recursive Contamination Loop

Here is where the empty report stops being an inconvenience and becomes an infrastructure risk.

In 2025 I led the drafting of a human-centric AI governance whitepaper for a community grants program, working with fifteen stakeholders across legal, protocol engineering, and civil society to write accountability guidelines for AI-driven DAOs. We secured $250,000 for pilot programs. The hardest clause in that document was not about kill switches or human override. It was about defaults.

Because an AI agent does not read a report the way a human does. It reads it as a retrieval document, embeds it, and inherits its shape. If a corpus of research contains thousands of documents whose structure is "nine dimensions, twenty-three tables, uniform N/A," then a retrieval-augmented agent asked to produce a new report will reproduce that structure — because structure is the easiest thing to imitate and the hardest thing to falsify. The agent is not hallucinating facts. It is hallucinating the form of rigor, which is more insidious, because form is what human reviewers skim.

This is the recursive contamination loop. Empty reports enter the corpus. Agents trained or retrieved against that corpus generate reports with the same shape. Those reports enter the corpus. Two cycles in, the statistical signature of "a thorough analysis" and "an analysis that found nothing because it looked at nothing" are identical in embedding space.

When I sat with the fifteen stakeholders on that whitepaper, the clause we fought hardest over was this: an agent must be permitted to emit "insufficient data" as a first-class output, and must never be permitted to emit a neutral default in place of an unknown. Because a neutral default is read as approval. A risk matrix with six green rows is a buy signal. A risk matrix with six N/A rows is, to a human under time pressure, also a buy signal — which is precisely the failure my founder walked into at two in the morning.

A Field Test: Five Signals

Since that morning I have been running a quick screen on every brief that crosses my desk, and I offer it here because it costs nothing and catches most of this.

One: a zero-length source list. If the document has a methodology section and no citations, the analysis is decoration.

Two: N/A density above forty percent in substantive cells. Some nulls are honest. A document that is more than four-tenths null is not an analysis; it is a request for information formatted as an answer.

Three: a call to action inside the deliverable. Real research does not end by asking the client to supply the inputs. When a "report" closes with a request to re-run the pipeline, you are holding the pipeline's error message, not its output.

Four: timestamps without block heights. A claim that something happened "on March 3" is unverifiable. A claim that something happened at block 21,884,102 is checkable in ninety seconds. Research that will not give you a block height is telling you what it is worth.

Five: risk matrices with no probabilities. Six rows, six categories, zero likelihood estimates, zero impact ranges. A risk matrix without numbers is a list of words that look like a matrix.

None of these five signals requires domain expertise. All five would have caught the document that cost my founder three days.

The Refusal Was the Alpha

Now the inversion, because the obvious reading of this story is wrong.

The obvious reading is that the pipeline failed. I want to argue the opposite: the pipeline's refusal to fabricate was the single most valuable output it produced, and the industry's inability to price that refusal is the actual crisis.

Consider what the alternative would have been. A pipeline that, given an empty fact list, filled the nine dimensions with plausible-sounding inferences. A technical section that guessed "likely an optimistic rollup" because the subject's name contained a familiar suffix. A tokenomics table with a supply figure reconstructed from a Twitter thread. A risk matrix with four green rows and two amber, because green and amber look more professional than N/A. That document would have been more useful to my founder and infinitely more dangerous. It would have converted an absence of information into a recommendation, and she would have sized a position on the strength of a guess wearing a suit.

The refusal is the product. It is the only part of the deliverable that was unambiguously true.

Which leads to a second inversion that I think is genuinely under-discussed. If N/A density is a reliable signal of disclosure failure at the level of a single issuer, then aggregated across a sector it becomes something else entirely: a disclosure thermometer. I have started tracking it. Across the 214 briefs I reviewed, the categories with the highest null density were not obscure micro-caps — they were mid-cap infrastructure projects with active foundations, public treasuries, and governance forums. The nulls clustered exactly where disclosure should be easiest: allocation schedules, multisig signer identities, sequencer operator sets, and upgrade timelocks. The information was not missing because it was hard to find. It was missing because the template asked the project for it and the project did not answer.

That is a screening metric nobody is publishing, and it is more informative than TVL. TVL tells you what capital is parked. Null density tells you what the issuer is willing to show you. In a consolidation market, where the whole game is separating durable protocol from durable narrative, I would rather hold the asset with a boring, fully disclosed three-of-five multisig and a published timelock than the one whose governance dimension reads "community-driven" in every brief on the market.

And here is the part that runs against every instinct the industry has trained into us: completeness is a red flag. A report with no nulls, no caveats, no "unable to verify" lines, and a clean green risk matrix is not a better report. It is a report that has stopped distinguishing between what it knows and what it was told. The most trustworthy documents I have reviewed in the past year are the ones that say, in plain language, here is what we could not confirm and here is why it matters. That is not weakness. That is the only form of strength that survives contact with a bear market.

Solidarity over speculation. The community that survives the next drawdown will not be the one with the most confident research. It will be the one whose research told the truth about its own limits, so that people could make decisions with their eyes open instead of their hopes up.

What We Owe the Next Cycle

I have spent twenty-seven years watching this industry mistake its own vocabulary for its own accomplishments. The vocabulary is now automated. That is the change, and it is not reversible.

So the question is not how to stop machine-generated research. The question is what we require of it. I think the answer is a schema change, and it is small enough to implement this quarter: every research document produced by a machine should carry a machine-readable confidence field, and every agent that consumes one should be required to propagate it. Not a disclaimer in the footer. A structured field, at the top, that says how many of the claims in this document trace to a third-party verified source, how many trace to the issuer's own statement, and how many are null. Any downstream agent that acts on the document without reading that field should be treated the way we treat a contract that ignores an oracle's staleness check.

The whitepaper work taught me that this is achievable — fifteen stakeholders with genuinely opposed interests agreed on it in six weeks, because the principle is not ideological. It is the same principle we accepted in smart contracts a decade ago: validate your inputs, fail loudly, never default to a value that looks like safety.

Culture on-chain, heart on-screen. We built an entire civilization of protocols on the premise that code should be verifiable by anyone who cares to look. It would be a strange irony if the research layer that explains those protocols became the one layer nobody can audit.

My founder did not size the position. She asked the project three questions instead — the allocation schedule, the multisig signers, the timelock — and got two answers and one silence. She is still waiting on the silence. That silence is the most valuable data point she has.

If a document can be rendered successfully from nothing, and a market can price that document as if it were something, and an agent can act on it before a human ever reads the nulls — then what, exactly, is the law here? And who is holding the conscience?