At 03:14 on a Tuesday, one of my screening scripts printed a green light on a lending market I had never heard of. Score 7.2 out of 10. Risk: medium. Recommendation: allocate.
The deconstruction underneath it was empty. No title. No source. No thesis. No information points. Nine analytical dimensions, zero facts. The pipeline coerced a null into a zero, the zero into a neutral prior, and the neutral prior into "medium risk." Then it drew borders around the whole thing and exported a PDF.
I killed the job at 03:22. By 09:00 two people had cited the PDF in a memo.
That's the part that should bother you. Not that a model hallucinated โ models do that constantly and everyone knows it. The problem is that a document containing no information looked structurally identical to a document containing information. Same headings. Same tables. Same confident little risk grades in the right-hand column.
I've run a cheaper version of this experiment with my own money. In early 2025 I gave an autonomous agent $100,000 and 50 trades of rope on Ethereum L2s, letting it chase social volume spikes in meme coins. It lost $30,000 in two weeks. The failure wasn't sentiment modeling โ the agent was actually decent at reading social velocity. The failure was that when its data feed dropped a field, holder concentration or liquidity lock status or whatever else, the null got imputed as "neutral," and neutral meant "tradeable." Thirty grand to learn that missing data is not the same as benign data.
Every research desk in this industry runs some version of the same nine-dimension template. Technical architecture. Token economics. Market structure. Ecosystem position. Regulatory posture. Team and governance. Risk matrix. Narrative and expectations. Supply-chain transmission. I've filled out hundreds of these. They aren't stupid โ the checklist is genuinely good. It catches the things people forget to look at.
The problem is the output contract. The template demands a value in every cell. Analysts are paid to deliver complete documents. Funds want something to circulate. So when the facts aren't there, the cell gets filled anyway โ with N/A, with "unclear," with a hedge, with a rating.
An empty field is a fact. An N/A is an opinion. They get filed in the same place and read the same way.
I learned the shape of this on the ETF side. In 2024, ETF approval wasn't a thesis โ it was a data event. The trade wasn't the headline; it was knowing exactly which filings were timestamped and which desks were still pricing a stale NAV against them. The edge lived in the cells everyone else had left blank because the paperwork was tedious. That's where arbitrage always lives.
Start with the coercion, because it's mechanical and it's everywhere.
In any competent data stack, None and 0 are different types. In research, they collapse into each other constantly. A missing APR is not a 0% APR. A missing TVL is not zero TVL. A missing unlock schedule is emphatically not "no unlocks." I have watched a token get flagged as fully diluted โ FDV/MCAP of 1.0, no upside โ because the circulating supply field was blank and the model defaulted to max supply. The same blank field, run through a different analyst's spreadsheet, produced the opposite conclusion: cheap, under-owned, asymmetric. Two contradictory trades, one empty cell.
Unlocks are where this gets expensive. A blank vesting row reads as "no scheduled supply pressure" in almost every summary I've seen, because the analyst has nothing to plot. Then the cliff ends and 14% of supply hits the book in a quarter and everyone acts surprised. The information was never hidden. It was absent from the report because the report had a slot and nobody filled it, and an unfilled slot in a table reads as a zero. The unlock calendar is the single most predictable sell-pressure event in this asset class, and it is routinely the emptiest row in the document.
Then there's the table itself. A risk matrix with six categories, all rated N/A, composited to "cannot be assessed" โ that's honest. But nobody reads the composite. They read the column. And a column of mid-weight adjectives anchors the human eye at "medium," because medium is where unlabeled things land. A table implies measurement even when it contains none. Prose at least has the decency to sound uncertain.
I got liquidated once on exactly this. A lending market I'd sized into had a collateral oracle with a 3,600-second heartbeat and a 0.5% deviation trigger. During a ninety-second wick on a thin venue, the feed printed a price that had already stopped existing. I was closed out at a number that lived for eleven seconds. The nine-dimension template had a row for oracle design and performance assumptions. It was blank in the report I read. Medium risk, the summary said.
The blank row was the entire risk. Everything else in that document was decoration.
Now the dimension where this does the most damage: users. Every ecosystem analysis wants DAU, MAU, retention. Fine. But retention against what baseline? I've been through dashboards where a stablecoin wallet cohort showed 60% ninety-day retention โ beautiful number โ and the cohort was almost entirely remittance corridors and inflation-hedge savings in currencies that had lost a third of their purchasing power in the same window. That's not product-market fit. That's a macro indicator wearing a growth metric's clothes. If your retention model can't separate survival-driven usage from speculative usage, your 30% health threshold is a number with no referent. It will tell you a payments product is thriving when what's thriving is the local currency's collapse.
Same disease, different organ: bridges. Cumulative cross-chain losses have cleared $2.5 billion and the industry still routes a meaningful share of TVL through them. In a risk matrix, "cross-chain exposure" defaults to medium because the analyst filling the cell doesn't have per-bridge incident data loaded โ verifier counts, wrapped-asset accounting, upgrade authority. So the single largest historical loss category in DeFi gets a mid-weight adjective and a footnote. I want the number in the cell, not the adjective. You don't rate a bridge "medium." You count its verifiers or you don't touch it.
The narrative dimension is the most honest of the nine and gets treated the worst. An expectation-gap table โ growth expected versus delivered, revenue expected versus delivered, delivery expected versus delivered โ with all three rows N/A, still produces a verdict. Usually "priced in." That is a conclusion about market psychology drawn from an absence of measurement. The only thing the table demonstrates is that nobody checked.
Transmission is where blanks compound. If your upstream column is empty โ no data on the infrastructure layer feeding the protocol โ you don't miss one cell, you miss the second-order trade entirely. Liquidity migrations, yield compression, composability risk: all of it propagates from the layer nobody loaded. And every cell has an expiry nobody stamps. Crypto information has a half-life measured in weeks, sometimes days if a governance proposal is live. A report with no per-cell timestamp isn't wrong, it's undated, which is worse, because it reads as current forever. I date every row I fill now. It looks pedantic. It has saved me twice.
Here's the uncomfortable part. The industry doesn't reward accuracy. It rewards coverage.
Alpha isn't the report you can produce. It's the report you can still defend six weeks later, after the unlock lands and the oracle drifts and the governance forum goes quiet. Nobody gets paid for the memo they refused to write. Nobody gets credit for the deal they killed in the first hour because the information-point list had three items on it and two were press releases.
So the incentive runs one direction: fill the cells, ship the deck, add a disclaimer. The blind spot isn't laziness. It's that completeness is legible and incompleteness is not. A report with holes gets interrogated. A report with nine sections and a risk grade gets approved.
The market doesn't pay for diligence anyway. It pays for being early and right, and those two are correlated far less than anyone admits.
Which means the actual edge in 2026 isn't generating research. Agents generate research now. Mine does, badly, at 03:14. The scarce skill is refusal โ the discipline to look at an input with no substantive information points and terminate the process instead of dressing it up. An absent unlock schedule isn't a gap in your model. It's the answer. An unnamed team isn't "N/A." It's a no.
Three rules I now enforce on every pipeline I touch. First: minimum viable input โ if the deconstruction returns fewer than five substantive information points, the report dies. No framework, no partial credit. Second: no null becomes a number, ever. Missing is missing, and it gets written in red. Third: timestamp every cell, because undated analysis is a liability that compounds.
The forward question is uglier than the mechanics. As agents take over the first pass of research across every desk in this market, the volume of complete-looking, information-free documents is about to go vertical. So: how many of the reports currently sitting on your desk would survive a null check โ and how much of your positioning is built on cells that were blank the whole time?