Last month, a governance contributor forwarded me a screenshot she could not explain. It was a finished project analysis — nine sections, forty-one tables, a five-star rating matrix, a color-coded risk register. Every cell was populated. Every conclusion was confident. At the top, the source field read, in a calm monospace font: Information points: none.
The pipeline had consumed nothing and produced everything. It took the research committee three days to notice.
I have been auditing crypto research since 2017, when I read more than fifty whitepapers for European startups at the height of the ICO mania. The lesson I carried out of that year is simple, and it has not aged: the most dangerous document in any investment memo is not the one with gaps — it is the one with no gaps at all. Confidence is cheap to manufacture and expensive to verify. This year, we handed the manufacturing to machines.
The convergence of AI and crypto was supposed to make due diligence better. In 2026, most serious DAOs run some version of an agentic research stack: one model extracts claims from a source, a second analyzes those claims across technical, tokenomic, market, and governance dimensions, and a third writes the memo. The architecture is elegant. It is also, as that screenshot proved, fragile in a very specific way.
Here is the thing nobody markets: an analysis pipeline is only as honest as its worst input contract. When the first stage returns a valid-looking structure with empty fields, the second stage does not crash. It does what it was trained to do — it completes the pattern. A blank table invites a number. A missing risk row invites a plausible risk. The model is not lying; it is doing its job in the absence of a job.
I watched this exact failure mode emerge in the DAO Literacy workshops I ran in Paris during DeFi Summer. We had built a voting interface that surfaced every proposal's key parameters in plain language. One quarter, a bug caused the requested-funding field to render as zero whenever the data feed lagged. Participation spiked. A dozen proposals passed. Nobody questioned the zeros, because zeros look like answers. We had built a system that was confident precisely where it was blind.
The empty template is not a bug in one model. It is a structural property of how we build agentic research, and it deserves the same scrutiny we give to a bridge contract.
Consider the two failure modes. The first is null propagation: the pipeline encounters missing data and passes the null forward, marking fields N/A and downgrading every rating to unrated. This is what the screenshot's engine actually did, and it is the correct behavior. The second is hallucination: the pipeline encounters missing data and fills it. In a bull market, the second mode is indistinguishable from competence, because invented optimism reads exactly like real optimism. The syntax is identical. Only the provenance differs, and provenance is invisible to a reader who is already in a hurry.
The reason the first mode is rare is economic. We reward the output, not the epistemics. A memo that says insufficient information to form a judgment is a memo a fund manager deletes. A memo that says strong technical fundamentals, undervalued relative to peers is a memo that gets forwarded. The model learns what we reward, and we have spent a decade rewarding fluency. The same gravity that pulled retail into projects with empty vesting schedules in 2017 now pulls research agents toward empty source fields. The shape of the trap never changed. Only the instrument did.
The fix is not a better prompt. It is a data contract with teeth. In the framework I helped design last year for verifying AI training-data ownership, we settled on a simple primitive: every claim entering an analysis must carry a verifiable credential naming its source, its timestamp, and the hash of the document it came from. If a claim has no credential, it cannot enter the pipeline. The second stage then operates on a typed set of information points, not on a hopeful blank slate. And critically, the pipeline carries a gate: if the number of credentialed information points falls below a threshold, the analysis aborts rather than completes.
That gate is what produced the empty template. It is the least glamorous line of code in the stack, and it is the one that saved the committee from publishing fiction.
There is a cryptographic argument here that the industry keeps skipping. We spent years learning that data availability is the binding constraint on rollups — that blob space is finite, that cheap blockspace gets consumed, that the cost of not accounting for data is a bill that arrives later, usually at the worst possible moment. Research pipelines obey the same physics. Every unverified claim is a blob you did not pay for, sitting in someone's memory, waiting to be priced. The difference is that rollup fees are visible on a dashboard and hallucinated fundamentals are not, until the position is underwater and the post-mortem begins.
I have written before that code is law, but people are the soul. The empty template is a rare case where the two agree. The code enforced honesty. The people, briefly, trusted it. And then the market did what markets do.
There is a second-order problem that almost nobody is modeling yet. DAO governance does not just consume research — it votes on it. When an agentic pipeline quietly fabricates a tokenomics section, that fabrication does not stay in a PDF. It becomes a talking point in a forum, then a slide in a delegate's deck, then a reason a quorum votes yes. I have sat in enough governance calls to know that a single confident table can move more votes than a week of open debate. The pipeline is not a research tool anymore. It is an opinion-injection mechanism with a publication schedule.
This is why the provenance layer matters more than the model layer. Attesting where a claim came from is a solved cryptographic problem. Attesting that a claim was never made at all is the hard part, and it is the part we keep outsourcing to vibes. A credential system that only signs real claims cannot distinguish a careful analyst from a fluent one. You need a signed record of absence — a verifiable statement that a search was performed, a source was queried, and nothing was found. Otherwise the honest pipeline and the lying pipeline produce identical artifacts, and the market will price them identically.
I learned that lesson the hard way in 2017. I had audited a decentralized exchange that promised instant settlement and, on inspection, had no working zero-knowledge proof underneath the promise. I could have sold that finding to a fund and moved on. Instead I published a guide I called The Ethics of Empty Vests, warning retail readers that a project's confidence is not evidence of its substance. It cost me a job. It bought me something more durable, which is the habit of asking what a document is hiding by being so complete.
Here is the counter-intuitive part, and it is the part that should worry anyone building in this space.
The committee did not flag the empty template as a success. It flagged it as a failure. The pipeline had produced no analysis, and the contributor who ran it was asked to re-run it until it worked. The guardrail that prevented a fabricated investment thesis was treated as a defect, because the deliverable was a memo and the memo was empty. In the same week, a competitor's agent shipped a glowing write-up of a project whose contracts had not yet been deployed. It was well-received. It was also, technically, fiction — but it had no empty cells, so it passed review.
The market does not pay for epistemic honesty. It pays for the appearance of rigor at the speed of automation. This is the same dynamic I watched in the NFT cycle, when projects with genuine cultural depth were outperformed by collections with better marketing calendars, and the same dynamic I watched during the Terra collapse, when the most rigorous risk analysis was the least read document on the forum. The uncomfortable conclusion is that the empty template is not a technical problem we can engineer away. It is a governance problem. Don't govern the exit; govern the entrance. If we want honest research, the credential has to be a prerequisite for the pipeline to run at all, and no credible sources found has to become an acceptable, even honorable, result. That is a cultural change, not a software release.
The empty template will not stay empty for long. Someone will ship a version that fills every cell beautifully, and it will win the quarter. The question for the next cycle is whether we will build the infrastructure to tell the difference before the difference costs us a fund — or whether we keep rewarding the model that hallucinates most fluently. Code can enforce the gate. Only people will decide whether to walk through it.