Last week a research pipeline handed me a nine-section protocol analysis. Every heading was populated. Every table was complete. Every row read N/A.
Zero information points. Zero project identified. Zero thesis. The document still shipped a risk matrix, a token supply breakdown, a Howey-test table, four star ratings, a disclaimer, and a prioritized remediation list. By formatting standards it was the most complete report I read this quarter. By information standards it was a null pointer rendered as a PDF.
I have stayed up auditing unverified fork code and found a mint function that printed infinite supply under a specific block height. I have watched a "decentralized" recovery plan hinge on one multisig key. I have never seen a failure this clean. Logic prevails where hype fails to compute โ but what does a pipeline return when it computes perfectly and the logic is simply absent?
Some background on how these systems are wired, because the wiring is the finding.
Machine-assembled crypto research runs in two stages. Stage one is extraction: parse a source document, pull the title, the information points, the core thesis, the projects named, the time sensitivity, the source quality. Stage two is analysis: take those extracted facts and run nine dimensions over them โ technical, tokenomics, market, ecosystem, regulatory, team and governance, risk, narrative, supply-chain transmission.
The economics pushed everyone here. Coverage is the product. A desk that wants two hundred assets a week cannot hire two hundred analysts, so it buys a pipeline. In the current market the buyer intent is simpler than it was in 2021: holders do not want price targets, they want to know whether the thing they are holding is structurally sound or quietly bleeding LPs.
That demand is exactly what makes formatted output dangerous. A risk matrix looks like diligence. A Howey table looks like legal review. Star ratings look like judgment. All three can be generated from zero facts, and none of them carry a field that says how many facts they were generated from.
Then there is the aggregation layer, which is where nulls get promoted. A nine-section report is rarely read directly; it is ingested by a screener that converts each dimension into a 0-100 score. Ask what that coercion does to a null. In most pipelines the answer is nothing โ the field is dropped and the composite is computed from the remaining sections, which silently reweights the result. A report with zero technical data and zero tokenomics data still produces a clean composite, because the average of what survived is always a number. The emptiness gets laundered into a score, and the score gets laundered into a position.
Here is what actually happened in the pipeline I reviewed. The extractor returned empty. No title, no information points, no thesis, no tags. Stage two received the empty payload, checked its own output contract, found the instruction "if a field cannot be assessed, mark it N/A โ do not fabricate," and complied. Every cell became N/A. The system did not crash. The system did not lie. The system produced a document whose shape was indistinguishable from a real report.
That compliance is the story. Not the bug โ the decorator that wrapped the bug.
Let me get into mechanics, because "the output was empty" hides three separate failure classes.
Class one: schema-valid, semantically empty. Most pipelines validate types, not truth. A payload like {"core_thesis": "N/A", "info_points": []} passes float, bool, and object tests with perfect marks. The string "N/A" is a valid string. An empty array is a valid array. If your validator cannot distinguish "a thesis exists and it says N/A" from "no thesis was extracted," your validator is decorative. In the sandbox I built for AI agents generating transaction payloads, I learned the same lesson in a more expensive medium: type-correct and intent-correct are orthogonal properties. A payload can be perfectly well-formed and mean nothing.
Class two: hard constraints beat soft constraints. Give a model nine mandatory sections and a soft instruction not to hallucinate, and you have created a pressure gradient. Under tension, the enforced constraint wins. That is why so much auto-generated research is shaped like analysis and weighted like noise. The pipeline I reviewed avoided confabulation for one reason only: the input was visibly empty. There was nothing to confabulate from. Now change one variable. Give stage two a source with one information point out of the twelve it expected. Nothing is visibly empty. Three sections can be written from that one fact. The other six get filled with priors so fluent that no reviewer flags them. That is the case that reaches a reader, and that is the case the N/A gate never fires on.
Class three: no provenance on numbers. A fabricated risk rating and a real one render identically. Unless every numeric in the output carries an offset back into the source text, the reader cannot separate a derived figure from a generated one. I spent three months in 2020 building a Python harness that replayed five thousand mock transactions between Aave v1 and Compound to measure oracle latency during volatility spikes. I found a four-second window. That number was defensible because I could point at the block. Take away the trace and you have a dashboard.
Consider the adversarial version, because it is easier than it sounds. Injecting a truncated source is cheaper than injecting a false one. You do not need to control a document's contents โ you only need to control how much of it arrives. Cut the fetch at forty percent, and the extractor still returns three or four information points, enough to pass any cardinality gate set low enough to be practical. The downstream report is now built on a biased subset, correctly cited, confidently formatted, and wrong in a way no reviewer can see, because every individual sentence traces back to a real line in a real document. That is a data-integrity attack, and it looks exactly like a Tuesday.
Prompt-auditing, in my experience, is ten percent injection and ninety percent input integrity. The logic bombs I catalogued in that agent sandbox almost never required a malicious instruction. They required a missing one.
The assertion layer is not exotic. It is six checks: hash and byte-length the source before extraction; assert a minimum information-point cardinality; ban "N/A" as a value and force an explicit insufficient_data object listing what is missing; require a source offset on every number; compute the confidence field as a function of source, never of fluency; and put a kill switch on the orchestrator so a zero-input run halts instead of rendering.
The kill switch matters more than the other five. A pipeline with one extractor and no assertion layer has the same structural flaw I documented on Terra Classic after the 2022 crash: a governance system that markets distributed decision-making and routes its emergency pause through a single key. One node, no quorum, total authority. Every research pipeline I have audited this year with a single extraction stage has the same shape โ a single point of failure wearing a nine-section costume.
There is a real trade-off and I will not pretend otherwise. Strict gates cost throughput. A cardinality assertion at three information points rejects sources that are thin but genuine. Teams route around friction; they add a force_output flag within a quarter, and the flag becomes the default. The engineering answer is not tighter gates alone โ it is making refusal a first-class artifact. A REFUSED: INSUFFICIENT_SOURCE object is a different data type from a report. It cannot be exported to the same dashboard, forwarded to the same client, or formatted into the same nine headings. Once refusal has its own schema, the pressure to decorate disappears, because there is nothing to decorate.
The counter-intuitive part: the empty analysis was the highest-integrity output in the entire system.
Most AI-safety attention in crypto is aimed at models inventing alpha โ a fake yield claim, a made-up partnership, an inflated TVL. Those are loud failures and they get caught, because someone always knows the real number. The quiet failure is a model inventing structure. Nobody catches structure. Structure is what we use to decide what to trust.
We already ran this experiment once. Between 2017 and 2021 the market learned to distrust the whitepaper: the roadmap slide, the advisory board, the token narrative. We replaced it with the dashboard, the audit badge, the risk matrix โ artifacts that look like evidence because they are formatted like evidence. A nine-section report built on zero information points is the same object wearing a better suit. The 2017 project I audited had a marketing site and an integer overflow; the 2026 pipeline has a risk matrix and an empty array. Different decade, same structural defect: the presentation layer is doing work the data layer never did.
There is a second inversion. The industry assumes more data fixes this. It does not. More sources, more feeds, more stages โ every added stage is another place where a null can be wrapped in a heading. Scaling coverage without scaling assertions just produces formatted nothing at higher frequency, and higher frequency is what makes it look like consensus.
And treat the empty payload as an attack surface, not an accident. The cheapest adversarial prompt injection is not a hidden instruction telling an agent to drain a wallet. It is nothing at all โ a truncated document, a failed fetch, a 200 response with a zero-length body โ routed into a system contractually obligated to produce nine sections. You do not need to break the model. You need to starve it and let it fill.
Vulnerability forecast: within the year, the serious research shops will publish their information-point counts alongside their conclusions, and "how many facts is this report built on" will become a field you can filter on. The rest will keep shipping formatted nulls, because a formatted null is easier to sell than a refusal. Logic prevails where hype fails to compute.
One question to run against everything you read this week, including this: how many information points was it actually built from? If the author cannot answer that, the document is a shape. And shapes do not hold value in a bear market.