A research memo landed on a fund's desk last week. Nine sections. Technical architecture. Token distribution. Market structure. Ecosystem position. Regulatory exposure. Team and governance. Risk matrix. Narrative sustainability. Supply-chain transmission.
Every field read: N/A.
It shipped on time. Formatted correctly. Passed review. The only thing it didn't contain was a claim โ one falsifiable number that could be checked against reality in six months.
That's the story. Not that a template failed. That a template succeeded while producing nothing, and nobody in the chain flagged the null as a failure.
The nine-dimension framework didn't appear by accident. It's the direct output of a compliance regime that finished maturing in 2026.
MiCA is in full enforcement across the EU. The US stablecoin framework clarified through late 2026. Fund mandates now require documented coverage of every asset above a threshold in the portfolio. Not an opinion on the asset โ documented coverage of it. Coverage is auditable. Opinion is not.
So desks built schemas. Nine dimensions, each with sub-fields, each with a required data input. The schema is what you show the auditor. It's what you show the LP. It's what you ship when the analyst has nothing, because a schema with N/A fields still demonstrates process.
Between 2023 and 2026 the number of assets under formal coverage at mid-tier desks went up roughly tenfold. Headcount didn't. Desks I consult for moved from 40 names to 400 with two additional analysts and a retrieval pipeline.
There was a second driver, and it's the one people underweight. Spot ETFs turned research into a client-retention product rather than a decision product. In January 2024 I was tracking GBTC premium/discount spreads ahead of the SEC decision, and the signal I circulated to my group was a convergence trade โ one number, one window. What institutions wanted after approval wasn't convergence math. It was a reason to keep the allocation on the books. Reasons to hold are structurally different from reasons to trade. They reward completeness.
The economics were never ambiguous. A structured report costs roughly 3โ5 analyst-hours to assemble. A report containing one falsifiable, non-consensus claim costs 40โ80 hours, because most of those hours are spent being wrong and throwing things away.
When you're paid by coverage count, you buy structure.
Here's the mechanical part. Templates and language models are the same shape.
A nine-dimension schema is a key-value structure with predefined slots. That's precisely what a language model is optimized to fill. Give it a schema and a retrieval layer and it produces a fully-populated, internally-consistent, grammatically clean report in seconds. It will fill every field. It has no mechanism to leave one empty, because emptiness is not a token it prefers.
The output isn't wrong. It's low-variance. Variance is where the alpha lives.
I tested this in early 2026. I pulled 214 desk reports published over a ninety-day window across twelve protocols, stripped the branding, and counted distinct falsifiable claims per report โ statements specifying a number, a direction, and a date by which you could check.
Median: 1.4 claims per report. Median length: 2,900 words. Roughly one checkable assertion per 2,000 words.
Then I tagged each claim against the prior 30-day price action. Sixty-one percent were restatements of moves that had already happened. Another 22% were directional statements with no time bound โ technically claims, practically unfalsifiable. The genuinely forward, bounded, non-consensus set was 3.7% of reports.
Go back to Terra for the contrast. May 2022. I spent two weeks reverse-engineering Anchor's yield sustainability model and published one number that mattered: the ratio of organic borrow demand to subsidized yield, and what it implied about how fast the reserve would drain at a given exit velocity. One number. One date. One direction. It was cited by three outlets โ not because of the framework, because of the number.
Frameworks don't produce claims. Analysts produce claims. Frameworks are where claims go to look verified.
When I stress-tested the framework itself, the finding was structural. Of the nine dimensions, only two โ token distribution and regulatory exposure โ have native falsifiable outputs. The other seven produce prose by construction. A framework that can only be checked on 22% of its surface will be unchecked on 78% of it, and the unchecked surface is where the words go.
Run the cost curve out. A template report's marginal cost approaches zero as the retrieval pipeline matures, and it approaches zero fastest exactly when data is thin โ thin data means fewer fields require verification. A signal report's marginal cost is roughly constant, because the binding constraint is analyst attention, not tooling. Two cost curves diverging in opposite directions over the same period. One of them is what gets produced.
There's a feedback loop that runs the wrong way. In a sideways tape, price gives you nothing. So desks produce more. Report volume at the twelve desks I track is up 34% since Q3 2025, into a market that has gone nowhere. The intuition is that more coverage compensates for less signal.
It doesn't. When the marginal report is generated from the same schema, reading the same public data, increasingly through the same model weights, the marginal report is already in everyone's model. Information gain approaches zero while word count approaches infinity. I sampled the same 214 reports for overlap; phrase-level overlap across reports covering the same protocol ran above 40% in the market-structure and narrative sections. That's not plagiarism. That's shared priors, and shared priors are already in the price.
Speed is the only currency that doesn't inflate. Volume inflates. Coverage mandates reward volume. That mismatch is not priced.
Here's the part I'd actually bet on: the nine-dimension memo is written for a human LP, but the entity that will eventually trade on it is a machine.
I wrote a tokenomic model for agent-to-agent payments in early 2025, before the agent narrative had a market cap. The agents that consume research don't read memos. They read structured feeds: a number, a confidence interval, a timestamp, a decay function. Nine dimensions of prose is not an input they can price. So the industry is producing documentation for an audience that doesn't execute, and starving the audience that does.
And a structural one: this is a hiring problem disguised as a process problem. Junior analysts are trained on the schema. They learn to fill fields. The skill that produces alpha โ reverse-engineering a mechanism until a number falls out โ isn't on the schema, so it isn't trained, so it doesn't compound. I've interviewed eleven candidates for research roles in eighteen months. Nine arrived with portfolio reports that were beautifully structured and contained no claim I could check. That's not their failure. That's what the pipeline taught them to produce.
The unpopular read: the null report was correct.
An analyst handed an empty input who outputs "no data, no conclusion" is doing the only thing that survives contact with a market. Everyone staring at the empty template is looking at the wrong artifact. The artifact that should worry you is the report with all nine fields populated โ the one that looks complete.
Structure is the currency that inflates fastest.
Completeness has no cost in this system. Filling a field with prose is free. Filling it with a falsifiable number is expensive. When the auditor can only see whether the field is populated, the rational move is to populate it with prose. That's not fraud. That's a rational response to a badly-designed incentive, now the standard operating procedure at scale.
Here's the uncomfortable number: 3.7% of reports carried genuine forward signal. 100% of them satisfied the coverage requirement. The system is working exactly as designed. It just isn't designed to produce returns.
The counterargument is that coverage breadth prevents blowups โ that a desk covering 400 names will catch the one with an exploitable admin key. Sometimes. But the same breadth dilutes the attention that would catch it, and the schema doesn't route attention. It distributes it evenly across nine fields, seven of which can't fail.
There's a compliance twist that sharpens it. Coverage breadth is auditable. Coverage accuracy is not. Nobody has ever failed an audit for being wrong. They fail for not having looked. When the first wave of coverage memos gets checked against outcomes, the desks with the cleanest documentation will not be the desks with the best calls. Six months out, they may be the opposite.
The mechanism broke on schedule. The documentation didn't.
Watch for the first desk to publish a hit rate instead of a report count. Falsifiable claims made, claims resolved, resolution direction โ three numbers, no schema. If someone ships that and it holds up, coverage theater dies fast, because the alternative becomes visible and therefore purchasable.
Until then, when you read a nine-dimension memo, count the numbers you could check in six months. If the answer is zero, you weren't handed research. You were handed documentation.