I was halfway through a flat white in a BGC cafe when a junior analyst slid a deck across the table. Nine sections, immaculate typography: Technical Architecture, Token Economics, Market Structure, Ecosystem Positioning, Regulatory Exposure, Team and Governance, Risk Surface, Narrative and Expectation, Value-Chain Transmission. Every heading followed by the same two characters — N/A.
He had pushed it through the research agent our boutique has been piloting since the spot Bitcoin ETF approvals. Stage one — document ingestion, the part that pulls a title, a source, a handful of tagged information points — failed silently. Stage two, the nine-dimension analysis framework, ran anyway. It couldn't admit there was nothing to chew on, so it produced a beautifully formatted framework for nothing at all.
We didn't lose the data. We lost the ability to notice the data was missing.
Most crypto research tooling in 2026 runs on a two-stage design, and I've watched it become the default inside boutique firms from Makati to Singapore. Stage one is ingestion and decomposition: scrape the source, verify the publisher, extract named entities, tag the discrete information points that every downstream claim is supposed to cite. Stage two is the evaluation layer — the nine dimensions above, each one meant to be populated from the tagged points.
The dependency is supposed to be hard. In practice it's decorative. The two stages are built by different vendors, shipped on different release cycles, and sold as a single subscription. Nobody owns the handshake. So when stage one returns an empty set, stage two doesn't halt — it improvises.
Why this is landing now and not three years ago comes down to who is buying research. The ETF era pushed something like ten billion dollars of institutional capital through the gate, and with it a procurement culture that wants artifacts. A memo. A score. A nine-tab framework with the firm's logo in the footer. Allocators don't buy insight, they buy evidence that diligence happened. That demand is real, and it is exactly the demand an empty shell can satisfy.
The economics here are uglier than the technology.
A research report is a social asset before it is an informational one. I learned this the expensive way back in 2021, when I bought three Bored Apes not for the metadata but for the room they unlocked — the launch parties, the introductions, the way a specific JPEG changed who returned your messages. The token was the ticket. The content was almost incidental.
Research decks work the same way. The nine dimensions are the ticket. They get you into the allocation call. And once a deliverable's value is social rather than informational, the market stops punishing empty content, because nobody in the room is checking citations while the deck is on screen. The failure isn't that the pipeline produced nothing. The failure is that producing nothing cost the pipeline nothing.
That's a fail-open design, and anyone who has spent a season inside DeFi knows the pattern. Oracle feeds are the clearest case. The industry spent years celebrating decentralization while most price feeds resolved through a handful of permissioned nodes — a centralization problem wearing a decentralization costume. I watched a lending market keep executing against a stale price for eleven minutes because the feed didn't revert, it just didn't update. The protocol did what it was told. The tell was never the wrong number. The tell was that the system had no branch for "the number is missing."
Crypto research pipelines have the same blind spot, and the nine-dimension framework makes it worse by dressing omission in the costume of rigor. Enumerate the dimensions out loud and you sound like you did work. Technical architecture. Token economics. Market structure. Ecosystem positioning. Regulatory exposure. Team and governance. Risk surface. Narrative. Value-chain transmission. It's a genuinely good checklist — I use a version of it myself when I sit down with a project in Manila.
But a checklist applied to a vacuum doesn't produce silence. It produces prose. And here is the mechanical reason: language models are trained to continue, not to refuse. Give one a heading and it will fill the heading. Give it nine headings and an empty evidence layer and it will fill all nine, fluently, with the register of competence and zero provenance. The output isn't a lie in any intentional sense. It's a continuation. It's the most probable next token, and the most probable next token after "Regulatory Exposure:" is never "insufficient data."
That distinction matters more in crypto than anywhere else I've worked, because our asset class prices narrative first. I've believed that since a sweaty night in Makati in 2017 when I dropped fifty thousand pesos into Icon and Waves off the back of a stage pitch and nothing resembling a model — and sold two hundred percent higher because the crowd agreed with the pitch. Sentiment led. Fundamentals never got the chance.
So think about what a hallucinated research report actually is. It isn't a neutral document sitting in a filing cabinet. It's a narrative input. It gets screenshotted. It gets quoted in a Telegram group in Cebu. It becomes the citation for the next thread, which becomes the reason a retail trader sizes up. Fake analysis doesn't fail to move markets. It moves them, and then it becomes the thing that retroactively justifies the move.
I've already seen the grassroots version. A six-page "alpha report" circulating in a local trading group last month — nine headings, confident language, a token I'd never heard of, and not a single number in it that could be traced to a source. Two hundred people read it. Some of them bought. That's stage two of a broken pipeline, running at street level, with real pesos attached.
The remediation paths from that shell report are worth reading closely, because they contain the actual design principle. Fix the ingestion layer and re-run. Supply the raw text and let the pipeline re-decompose. Or, if the source is genuinely gone, degrade — but degrade to a hard floor: a title, a publisher, three to five core information points, and nothing else. The analysis layer must never be allowed to outrun the evidence layer. Minimum viable fields, not maximum viable prose.
The same document demanded confidence tiering — every conclusion labeled as explicit in the source, reasonable inference, or high speculation. That's the right instinct, and it's also the tell for why this keeps happening: in a true information vacuum, everything collapses into tier three, and nothing in the research business is set up to print "high speculation" on nine consecutive slides. The labeling discipline breaks exactly when it's needed most.
Here's where I'll push back on my own industry.
The empty shell isn't the scandal. The full shell is the scandal. A pipeline that fails loudly and hands you nine N/A sections is annoying. A pipeline that fails quietly and hands you nine confident sections built on two scraped paragraphs is dangerous, and it is infinitely more common — because thin sources plus fluent prose is the default state of automated research, not the exception. That N/A deck was the first honest thing that agent produced all quarter.
And "more data" is the wrong fix. Scraping harder doesn't help if the analysis layer has no branch for absence. The fix is cultural before it's technical: make "we don't know" a payable deliverable. Right now, in a tape that only goes up, confirmation is the product. Nobody writes a check for a report that says the evidence doesn't support a view — which means the market is actively selecting for pipelines that hallucinate.
The DYOR disclaimer at the bottom of every one of these reports isn't ethics. It's a liability transfer. The producer keeps the fee; the reader absorbs the epistemic risk. That trade only looks fair because we've all agreed not to read the fine print.
So here's what I'm watching as the next institutional wave writes its checks. Not whether the pipelines get smarter. Whether they learn to fail closed — whether stage two can look at an empty evidence layer and refuse to speak, the way a decent oracle rejects a stale print instead of executing against it. The technology to do that is trivial. The incentive to do it is not.
When the next nine-section deck lands in front of you, ask one question before you read a word of it: could this document have been produced if the pipeline had found nothing? If the answer is yes, you're not reading research. You're reading a ticket.