Nine dimensions. Technical architecture. Token supply and unlock schedule. Market microstructure. Ecosystem position. Regulatory exposure. Team and governance. Risk matrix. Narrative sustainability. Supply-chain transmission.
Every single cell came back the same: N/A โ insufficient information.
I read maybe forty crypto research reports in the last six weeks. Bull market, so the volume is up. Every fund, every KOL, every paid Telegram group is shipping "deep dives" at machine speed. Thirty-nine of them had a number in every box. Clean tables. Confident Howey analyses. Unlock schedules mapped to the month.
One of them had nothing. It was the only one I trusted.
The document was a two-stage research pipeline. Stage one takes a source article and decomposes it into atomic information points โ the smallest analyzable units, each tagged with provenance, type, and content. Stage two runs nine analytical dimensions on top of those points: technical, tokenomics, market, ecosystem, regulatory, team and governance, risk, narrative, supply-chain transmission. Each dimension must produce at least three conclusions and two inferential notes, every one of them citing a specific information point, every inference graded by confidence.
Then stage one returned an empty template. Title: blank. Information points: zero. Projects identified: none. Time sensitivity: unassessed.
The pipeline broke. And the second stage did the one thing almost no crypto research product does in 2026 โ it refused to move. It labeled every field unassessable, declared that any "depth" produced from zero inputs would be pure hallucination, and spent the remainder of the document specifying how to repair the upstream feed.
That refusal is the story.
Here is the mechanical problem. A large language model asked to fill a Howey test table will fill it. Four elements โ money invested, common enterprise, expectation of profit, efforts of others. There is no empty state in the output distribution. Give the model a token ticker and no documents, and it will interpolate from priors. And crypto priors are contaminated. The training corpus is saturated with marketing copy, launch announcements, and partnership press releases that were never true. A fabricated unlock schedule is typographically indistinguishable from a verified one.
I have audited enough of these to know the tell. In late 2017 I spent six weeks on the 0x Protocol v2 contracts โ not the whitepaper, the GitHub. I found three re-entrancy vectors and submitted them publicly. I refused to sell until the patches landed. That trade paid because verification was expensive and almost nobody was paying it. Nothing about 2026 has changed the arithmetic. Generation is free now. Verification still costs six weeks.
The framework's design encodes exactly that. Every conclusion carries an evidence citation. Confidence is graded high, medium, or low. Hallucination is named explicitly as the failure mode being defended against โ not as an abstract risk, but as a specific mechanism: plausible output generated in the absence of grounding. Most research products ship the generator and skip the filter.
The risk section is the cleanest example. Six categories โ technical, market, operational, regulatory, competitive, narrative โ and every one returned an unassessable rating. Most reports I read rate everything "medium." That is not risk analysis. That is a refusal to be wrong dressed as caution. An all-empty matrix is a different and far more useful statement: there was no risk carrier in the input to attach a rating to.
Look at what the remediation path actually asks for. Step one: confirm the source text was passed in. Step two: re-run decomposition until you hold five to ten information points with provenance. Step three: if information density is still low, supplement with secondary sources โ DefiLlama for TVL, Dune for on-chain flows, Etherscan for contract state, audit reports from Trail of Bits, OpenZeppelin, or CertiK. That is a provenance specification. That is what a research pipeline is supposed to look like before it is allowed to say anything.
The report also chose a dual-track output: the full schema with placeholders, plus an executable repair path. That is the correct engineering response to a null input โ surface the contract, then specify the fix. It is also, incidentally, exactly how you should read any AI-assisted research you are paying for.
Code doesn't care about your feelings. Neither does a blank field. The nine dimensions are a checklist, and a checklist is worth precisely the evidence sitting behind it. Strip the evidence and you have a template โ the same template that gets repackaged and resold as institutional-grade analysis to people who will size a position on it.
The consensus right now is that autonomous agents will replace the crypto analyst. Wrong framing. Agents will replace the writing. They will not replace the verification, because verification requires touching state that is not in the model โ the contract, the multisig, the treasury wallet, the actual reserve attestation.
And the industry has a structural incentive to keep that gap open. Fabricated depth sells. A table with a number in every cell closes a funding round; a table full of nulls does not. Which is why "liquidity fragmentation" keeps getting relaunched as a problem only a new token can solve, and why cross-chain bridges โ cumulative losses north of $2.5 billion โ still sit in the critical path of half the yield strategies I review. Yield is the bait, rug is the hook. The marketing is downstream of the mechanism, and the mechanism does not read the marketing.
Panic sells, liquidity buys. Right now the panic is FOMO, which is the same trade with better lighting. Bull markets are precisely when the empty template is most valuable, because that is when nobody wants to see one.
Track the null ratio. Not word count, not dimensions covered โ the proportion of cells your research provider is willing to leave blank. A team returning 15% null is telling you where its evidence stops. A team returning zero is telling you something else entirely.
Ask them directly: what did you refuse to answer this quarter, and why?
If the answer is nothing, you already know what you are holding.