Last month a fund I advise forwarded a due-diligence packet for a project that had raised $100 million in eleven days. Twenty-two pages. Nine analytical dimensions. Technical architecture, token economics, market structure, ecosystem position, regulatory posture, team and governance, risk matrix, narrative sustainability, supply-chain transmission. Charts. Footnotes. A disclaimer.
Every single field read N/A — information insufficient.
Not one field. All of them. Someone had taken a rigorous analytical framework, fed it a null struct, and formatted the refusal into a deliverable. The document was generated in eleven seconds. My second read of it took four minutes.
I have spent nine years reading documents like this. Usually the failure runs the other way — a report stuffed with numbers that do not reconcile. This one was honest. It refused to fill the blanks. That refusal, in the current market, is the rarest artifact I have encountered this quarter.
The code does not lie; only the auditors do.
Why the shape of research changed before the content did
Bull markets do not produce more research. They produce more research-shaped objects.
In 2017 I spent six weeks reverse-engineering the token contracts of Ethereum Gold. The deliverable was nine pages. Four of them were function-level analysis of the minting path, where I had found an integer overflow that would let any holder manufacture supply. The team raised $12 million anyway. Two weeks after launch the exploit fired and the treasury drained. My report had one chart. It was ugly. It was correct.
That economics no longer closes. A paid research group in 2026 ships 400 pages a week. An AI research agent ships 4,000. The marginal cost of a plausible nine-dimension framework has collapsed to roughly the cost of the inference call. What has not collapsed is the cost of verification — one block explorer, one mempool, one ledger, read by a human who knows what to look for.
That asymmetry is the entire story of this cycle. Supply of narrative: infinite. Supply of verifiable facts: unchanged.
The framework in that packet was built around a concept called an information point — the smallest independently verifiable unit of fact. Who did what. Which parameter. What value. What timestamp. The rule is that every conclusion in all nine dimensions must trace back to at least one information point. It is a chain-of-custody constraint, and it exists for one reason: to make hallucination structurally impossible rather than merely discouraged.
When the information-point list is empty, the chain breaks at the first link. Everything downstream of that break is fabricated by construction. The framework failed closed. Most do not.
Four things the null struct proves
Report density is measurable, and the market median is near zero.
I built a small counter. It does not read prose and does not care about tone. It counts information points per 1,000 words: a 0x-prefixed address, a transaction hash, a block number, a named parameter bound to a value, a wallet cluster with a stated clustering method. Last quarter I ran it across 340 documents marketed as deep research by paid groups and AI research agents. Median density: 0.8. Mean: 3.1, dragged upward by four outliers that turned out, on inspection, to have been written by people who had actually opened a block explorer. My own audit notes run between 14 and 22.
The definition matters, so be strict. "TVL of $2.1 billion" is not an information point. "TVL of $2.1 billion at block 21,884,301, read from the protocol's own accounting contract, reconciled against DefiLlama within 1.4%" is one. "Strong community" is not an information point. "4,112 unique senders to the fee contract in 30 days, top 10 wallets 61% of volume" is one.
A document at 0.8 density does not contain analysis. It contains formatting.
Attention is lumpy in real forensics and uniform in generated text.
Real investigations have an asymmetric shape. When I reconstructed Alameda's internal transfers in 2022, roughly 40% of the write-up covered eleven wallets and just over 500 transfers. Everything else — the market commentary, the timeline, the context — compressed into a few paragraphs. The bug lives in one place. The document follows the bug.
Generated reports are symmetric. Nine dimensions, nine equal sections, three bullets each, one table each, one risk rating each. That symmetry is the fingerprint, and it is cheap to test: