Last week I ran a controlled test on a research pipeline I built myself. I fed it a first-stage output containing zero information points. Every field was null. No title. No source. No project name. No claims, no timestamps, no extracted evidence. The upstream stage had failed silently β a broken field mapping between two nodes, the kind of plumbing failure that leaves no trace in a log dashboard.
The downstream stage returned nine complete analytical dimensions. A risk matrix with six categories. A five-star information-value rating. Three prioritized action items. A recommended remediation path. It even appended a professional disclaimer about not constituting investment advice.
Total input: nothing. Total output: a document indistinguishable in shape from every other document that pipeline had ever produced.
The ledger never lies, only the narrative hides. In this case the pipeline had invented a ledger to hide behind.
Context
Anyone who has spent the last eighteen months inside crypto data infrastructure has watched the same substitution happen. Human-authored research is being replaced upstream by agent-authored research, and the agents are increasingly allowed to act. In 2025 I led development of a verification protocol for AI-generated on-chain content β 200 agent behaviors integrated into Dune dashboards, tracking roughly $500 million in automated trading activity. The point of that project was never to prove AI can write. It was to find out how often AI writes something that is not true, and how quickly that something gets spent.
That second number is the one that matters now. A generated sentence is cheap. A generated position is not.
The industry has spent three years building pipelines: extract, summarize, score, publish. Research desks at three of the funds I advise now run some version of this stack. The stack is fast. It is also structurally biased toward output. Every stage in it is scored on whether it produced a result, not on whether the result was warranted by the input. That incentive is invisible until the input goes to zero β and then it becomes the only thing the pipeline is doing.
The economics explain the spread. Inference is priced per token, and a research desk buys throughput. A pipeline that abstains looks like a pipeline that failed to deliver, and a pipeline that always delivers looks like a cheaper research analyst. No vendor in this market has an incentive to advertise how often its system says nothing, so none of them measure it. The metric does not exist because the market has not asked for it.
Core
I ran a null-input test across seven research pipelines: the one I maintain, plus six open-source agent frameworks whose prompts and schemas are publicly documented. Same payload β a first-stage object with every field nulled and a schema validator confirming structural validity. Then I measured two things: whether a report was produced, and whether any claim in that report was traceable to the input.
Six of seven produced a full-format report. All six filled 100 percent of their required fields. Zero percent of their claims were traceable to the input. The seventh pipeline halted, because someone had written an explicit gate: if the evidence list is empty, return nothing and log the failure.
That gate is three lines of code. Six teams had not written it.
The failure mechanism is not mysterious. It has three parts, and all three are design choices, not model defects.
1. Schema coercion. A structured-output pipeline does not have a silence option. It has a set of required keys. When the evidence that should populate those keys is absent, the model does not leave them empty β it fills them. An empty field looks broken. A filled field looks finished. Under a schema with no null-tolerant state, fabrication is the path of least resistance.
2. Refusal scored as failure. I reviewed the evaluation harnesses behind four of the six pipelines. In three of them, an abstention β a response that declines to conclude β was counted as a failed task. The rubric rewarded completeness. Nothing rewarded correctness. A system trained on that rubric learns, correctly, that guessing is cheaper than saying nothing.
3. Format-completeness instructions. This is the one I want readers to sit with, because it is the most common and the most defensible-sounding. Instructions like output must be format-complete, fill every field, always provide a conclusion are written by people who want reliable parsing downstream. They are also, mechanically, instructions to fabricate whenever the input is thin.
The result is a class of document I have started calling a confident null β a report whose internal structure is flawless and whose external referent does not exist. It passes schema validation. It passes length checks. It passes tone checks. It fails the only test that matters, which is whether any number in it can be walked back to a source row. A number without a source row is a rumor.
What gets fabricated first is predictable. In the six pipelines that produced output, the fields most reliably invented were risk severity ratings, categorical labels as compliance status or sequencer quality, forward-looking timeline estimates, and anything requiring a confidence-style qualifier. Those are precisely the fields a consuming system keys on. The prose was generic and harmless. The taxonomies were not.
The inversion problem. Here is where a formatting defect becomes a market defect. Downstream consumers of these reports β dashboards, alerting systems, and increasingly automated execution β do not have a representation for unassessable. They have green, amber, and red. When a risk engine returns unable to assess, the consuming system maps it. It maps it to neutral, or to the previous value, or to green.
I have seen this exact inversion before, at scale. In 2022, mapping liquidity holes across Aave and Compound after the Terra collapse, I found that roughly 30 percent of the risky positions in my sample were not flagged as risky β they were flagged as unreadable. Stale oracle reads, missing collateral fields, ambiguous debt tokens. The positions my model could not assess were, in aggregate, the most dangerous positions on the book. Inability to assess was not a neutral signal. It was the signal.
An empty first-stage output is not a statement that a protocol is safe. It is a statement that you know nothing about it. Those two readings have opposite implications, and pipelines that map null to neutral systematically produce the wrong one.
Tracing the ghost liquidity back to its source
Every fabricated analysis has a downstream cost, and it is usually routed through something that looks automated and therefore credible. A null-input report that rates a stablecoin issuer compliance risk: moderate becomes a risk-parity weight. A null-input report that rates a rollup sequencer risk: low becomes collateral eligibility. Nobody audits the input. Everybody consumes the output.
I have spent seventeen years in this industry, and the pattern that recurs always has the same shape: an instrument that is trusted more than it is verified. In 2018 I audited 47 early-stage Ethereum contracts and found critical distribution vulnerabilities in 12 of them, and the reason those 12 were dangerous was never that the code was exotic β it was that nobody had read it, and everyone assumed someone had.
Contrarian
The reflexive conclusion here is that AI models hallucinate and therefore should not be trusted with research. That conclusion is comfortable and it is wrong.
The models did what their constraints told them to do. The failure lived in the rubric, the schema, and the operator's tolerance for silence. I built one of those six pipelines, and the hallucination was mine before it was the model's. My own instinct β standardize the output, fix the skeleton, make every report comparable to every other report β is precisely the instinct that generates confident nulls. I have written in defense of prescriptive standardization for years. It is a real virtue. It also has a failure mode, and the failure mode is a perfectly formatted document about nothing.
The proposed fix, a human reviewer in the loop, does not close the gap either. A reviewer reads the output, not the input. Give a competent analyst a fluent nine-section report with a populated risk matrix and they will challenge the ratings, not the provenance. The only review that works happens one layer up, at the payload: what did the first stage actually return, and is that object non-empty and source-verified before anything downstream is allowed to run on it.
There is a second, subtler trap. Discovering that a pipeline fabricated a report on empty input does not license the opposite conclusion β that the underlying project is dangerous. Correlation is not causation, and null input is not evidence of misconduct. The correct output of an empty dataset is not a red flag and not a green flag. It is an abstention, with a timestamp and a reason code. Anything else is noise wearing the costume of signal.
Takeaway
Watch three things over the next quarter. First, whether research infrastructure vendors start shipping explicit input-validation gates and publishing how often those gates fire β an abstention rate, treated as a first-class metric rather than an embarrassment. Second, whether the agent dashboards now feeding institutional flow introduce a distinct fourth state for no data, rather than defaulting null to neutral. Third, whether the evaluation harnesses behind these systems stop scoring refusal as task failure.
None of that is technically hard. It is a schema change and a rubric change. What makes it slow is cultural: the industry treats a blank field as a bug and a confident paragraph as a feature, when on a null input the truth is exactly the reverse.
The uncomfortable version of the question is not about anyone else's stack. It is about the one you already trust. When was the last time your pipeline returned nothing β and when it did, did it tell you, or did it write you a report instead?
An audit that cannot fail is not an audit. A pipeline that cannot abstain is not research.
