Last month I pulled the audit logs from a research desk I advise. Eleven contributors, human and otherwise. Over a 90-day window they published 1,240 deep dives across three verticals. I asked the pipeline one question: for how many of those pieces did the source-material parser return a non-null payload?
Thirty-one came back empty. Not thin. Empty. A CMS migration upstream of the scraper threw a 500, the retry logic swallowed it, and the downstream model received a null string where a source article should have been. The model did what models do. It produced a complete nine-dimension analysis anyway โ technical positioning, supply schedule, regulatory exposure, risk matrix, ecosystem map, the entire scaffold. Every field populated. Star ratings out of five. A disclaimer at the bottom, properly formatted.
Those reports averaged 2,900 words each. They were published. Some were syndicated. Two were quoted approvingly in a Telegram channel with 40,000 subscribers. Nobody caught them for six weeks, and the person who eventually did was not looking for them. I was auditing something else entirely. That is usually how it works.
On-chain analysis has one property no other financial research discipline can claim: its primary sources are public, permanent, and independently verifiable. I learned the value of that in 2017, at eighteen, tracing the Parity multisig exploit by hand across fourteen wallet clusters over four weekends, mapping dusting patterns until 60% of the stolen ETH resolved into three exchange deposit addresses. The lesson I took from it was that every digital footprint is permanent and decipherable. I have repeated that lesson in every briefing since.
The lesson has a second half, and the industry keeps skipping it. If every claim about a public chain is trivially falsifiable, then fabrication should be economically irrational here. It should be the one sector where lying doesn't pay. It isn't. Fabrication is thriving, and it is thriving specifically because the volume of output has decoupled from the volume of verification.
The mechanism is not mysterious. In 2026, an estimated 40% of DeFi lending activity is initiated by autonomous agents rather than humans โ that figure comes from my own work on Agent-to-Human Interaction Ratios, filtering transaction metadata for known AI wallet signatures. The commentary layer has followed the activity layer down the same path. Research desks now run model-assisted pipelines because the alternative is not publishing at all, and not publishing is indistinguishable from not existing.
Symmetry is the detail I keep returning to. If 40% of lending activity is machine-initiated, and a comparable share of the commentary describing that activity is machine-generated, then a meaningful fraction of this sector is now talking to itself. Agent A executes an arbitrage. Agent B observes the resulting volume spike, drafts a 1,200-word interpretation, and publishes it. Agent C, a portfolio agent at a small fund, ingests that interpretation as a sentiment input and adjusts a position. No human sits in the loop at any of the four steps, and every step is individually defensible. The failure compounds silently.
To understand how wide the verification gap has become, I ran a structured audit across fourteen desks and 3,100 reports over six months. For each report I matched every discrete claim against its cited source. I sorted them into three buckets: anchored, where the claim traced cleanly to the source material; adjacent, where the claim pointed at a real on-chain object with no causal relationship to the assertion; and orphaned, where nothing existed at the other end. I also built a structural metric โ the ratio of populated template fields to source bytes, which I call the completeness ratio.
Real analysis is ragged. This is the first thing you learn in forensics and the last thing you learn to accept. During the Parity trace I spent three weeks unable to attribute a single hop in the laundering chain, and the write-up said so, in a sentence that made me look incompetent to the two people who read it. That sentence was the most honest thing in the document. Honest desks show a curve: completeness rises with input size and plateaus well below 1.0. Confabulating pipelines do not have a curve. They cluster at a completeness ratio of roughly 1.0 regardless of input size โ full template, zero source, full confidence. The empty-input cluster sat at the top of that distribution with a mean completeness of 0.97.
The honest desks in my sample averaged an N/A ratio of 0.11 โ one field in nine explicitly marked as unknown. The empty-input cluster averaged 0.02. Confabulated analysis is not merely more complete than real analysis. It is more complete in exact proportion to how little it had to work with.
Here is the part that should worry anyone allocating capital against this research. Of the 340 discrete claims I hand-verified across those 31 reports, only 9% were outright fabricated โ addresses and transaction hashes that resolve to nothing on any explorer. That is the failure mode everyone fears, and it is the least dangerous one, because a dead hash fails a block explorer check in three seconds. The real damage sat in the adjacent bucket: 61% of claims anchored to genuine, verifiable, entirely irrelevant on-chain objects.
One report stated that a lending protocol "lost 40% of its LP base in seven days, indicating a governance-driven exit." The 40% figure was real. It belonged to a different protocol. The wallet cited was real too โ a treasury multisig that had executed zero transactions in that window. Two true facts, glued together by a false conjunction, formatted as a causal claim. Verifying either component individually returns a green check. The report was structurally unfalsifiable at the speed most readers operate.
Volume spikes don't lie. They simply refuse to explain themselves, and the distance between a spike and its cause is precisely where this kind of work gets done. The code doesn't lie. It also doesn't volunteer anything. We don't get to outsource the interpretation to the ledger, and we don't get to outsource the check to a model that has never seen the source.
I then pulled the funding layer, because research desks do not run on goodwill. Three of the fourteen are underwritten by "research DAOs" that allocate treasury to content production. I scraped the voter records on eleven funding proposals across those three organizations. Turnout was 4.1%, 3.7%, and 2.9% of eligible tokenholders. On each of the three, the top twelve wallets controlled between 18% and 26% of voting power. I have been running this measurement since 2020, when I scraped 5,000-plus Aave governance votes and found 15% of voting power concentrated in twelve entities. Six years later the surface area has grown by orders of magnitude and the concentration has not improved. The content that shapes retail positioning is underwritten by a governance process fewer than one in twenty holders participates in, and steered by a dozen wallets that have never published a correction.
Between the hash and the human, there is a silence. That silence is where fabrication lives. Machines generate the first draft; humans are supposed to occupy the silence โ the pause that opens the source field before publishing. In 26 of the 31 empty-input cases, no human opened the source field at all. The pipeline had a review step. The review step was a checkbox, and checkboxes get checked.
I want to flag a false lead, because I chased it for two days and it cost me a client meeting. All 31 reports appeared within a 72-hour window, shared identical section headers, and reproduced the same three typos in the disclaimer. That is a fingerprint. I read it as a coordinated campaign โ a single actor seeding a narrative across nine desks. It was not. It was one shared scraping vendor pushing a bad deploy to every customer at once. The correlation pointed at conspiracy; the mechanism was a misconfigured retry handler in a Frankfurt data center. The forensic instinct to find intent will mislead you whenever the underlying cause is banal, and in infrastructure the underlying cause is almost always banal.
The economics explain why this persists better than any accusation of malice. Producing one of those reports cost approximately $0.30 and four minutes of compute. Verifying one properly cost me about 90 minutes, and the deliverable was a paragraph explaining why the report was wrong โ a paragraph that earned a fraction of the engagement. Truth-telling in this sector is structurally unprofitable at the individual level. On-chain transparency does not fix it, because the lie is not on-chain. It lives in the inference layer, one hop above the data, where nothing is hashed and nothing is immutable.
The obvious conclusion โ that language models cannot do research โ is wrong, and I do not hold it. The models in question were not malfunctioning. They were executing an objective function that asked for a complete, publishable artifact, and they delivered one. Nothing in that pipeline assigned a cost to the absence of data. This disease also predates transformers by decades: sell-side equity research ran the same template for fifty years, forty pages, every section filled, because a blank section is a career liability and an invented section is not. The difference is throughput. A junior analyst could produce four bad reports a week. A pipeline produces four hundred, and the ones with the highest completeness ratios score best on the engagement metrics that determine next quarter's budget. We have built a system that rewards the appearance of knowing and penalizes the accurate admission of ignorance, then expressed surprise when the ignorance is hidden.
The next signal to watch is not a price level. It is a control. I expect input-validity gates to appear in research pipelines the way circuit breakers appeared in exchanges โ a hard refusal to publish when the source payload is null, enforced upstream of the model rather than in a checklist downstream. Watch for source-coverage ratios and N/A density to show up on research dashboards next to slippage and latency. The desks that adopt them will look weaker for one quarter, because their reports will have holes in them, and then they will be the only ones worth reading. If your research desk cannot tell you what it does not know, what exactly are you paying it for?