Seven Null Fields: The Due Diligence Pipeline That Admitted It Had Nothing

Last Tuesday, at 06:40 Rome time, a research artifact landed in my inbox. It ran roughly seven hundred words and contained zero facts.

I want to be precise about that. Title field: empty. Source field: empty. Information point list — the array that is supposed to hold every claim, every piece of evidence, every verifiable datum — returned as an empty collection. Field after field, the same value. Not evaluated. Not classified. Absent.

One line deserves preservation. The source quality field instructed the reader to "judge based on the information point source field." The information point source field was empty.

That is not an oversight. That is a circular dependency wearing a lab coat. The math is perfect; the reality is broken. The template compiled. The analysis never existed.

The crypto research industry has spent two years industrializing output and calling it coverage.

The mechanism is simple. After the January 2024 ETF approvals, allocators who had never touched a private key suddenly needed documentation — something with headers, a rating, a PDF, a defensible signature at the bottom. Demand for the artifact outran demand for the finding. Vendors responded the way vendors always respond. They automated the wrapper and left the contents to chance.

By 2025 most mid-tier due diligence shops were running the same stack. A crawler pulls the project's own documentation. A language model summarizes it. A scoring template assigns weights. A human — increasingly junior, increasingly compensated per report rather than per correct call — glances at the output and ships it.

The economics explain the architecture. A genuine teardown of a DeFi protocol costs fifteen to forty analyst-hours. My LUNA memo in May 2022 consumed seventy-two hours of simulation across the Luna Foundation Guard's reserve composition before I was willing to commit a single conclusion to paper. A template-filling pipeline costs tokens. The ratio between the two is roughly one to a thousand. In a bear market, where fee revenue has contracted and headcount has been cut, that ratio does all the talking.

So the artifact I received is not an anomaly. It is a specimen. It is what the research function looks like when it has been optimized for delivery and never for verification.

Three defects, in ascending order of severity. None are unique to this pipeline. All are structural.

Defect one: the schema can be satisfied without content. A template is a type signature. analyze(article) → Report will typecheck whether Report contains eleven verified claims or eleven nulls. Solidity taught me this in 2021, at my own expense. I audited Rainbow Bank's staking contract before its thirty-million-dollar launch and found an integer overflow in the reward calculation. The function returned the correct type. The auditors had signed off because the code compiled and the types matched. Forty-eight hours after launch, twenty-eight million dollars left the contract through a path the type system considered valid. A structure that permits emptiness will eventually be filled with emptiness. The pipeline that produced my artifact had never been told which fields were fatal. So none of them were.

One more detail I have been unable to stop thinking about. The domain confidence field returned "not evaluated." The domain label field returned "unclassified." Both are terminal states. Nothing in between was recorded — no partial evaluation, no candidate labels, no reason for the stop. A state machine reached its final condition without emitting a transition log. Between the commit and the block lies the trap: the interval where work is supposed to happen is precisely the interval this pipeline left undocumented. I have seen the same pattern in contracts that skip validation because the happy path never triggers it. The check is not missing. The check is unreachable.

Defect two: circular provenance. The source-quality field deferred to the information-point source field. The information-point source field had nothing in it. This is an oracle reading its own feed — the exact failure mode I documented in 2026 while auditing an "autonomous" yield agent, where every trading decision traced back to a single founder's backend server. The agent reported confidence. The confidence was generated by the process that generated the trade. There was no independent observation anywhere in the loop. Trust is a variable that must be zero, and a provenance chain that terminates in itself evaluates to zero. It simply does so silently.

Defect three: silence has no price. The empty analysis was delivered. It was presumably billed. Nothing in the pipeline penalized it. Compare that to the MEV measurement I ran on Uniswap v3 in 2023. I bypassed the dashboard and read the mempool directly, and found that roughly 40% of what users paid on popular pairs was not fees but validator bribes, with about three dollars of every hundred reaching liquidity providers. The dashboard said the cost was a fee. The mempool said the cost was an extraction. The distance between those two readings is the entire distance between a pipeline that observes and a pipeline that transcribes.

The hidden cost ledger.

Cost of a real protocol teardown: 15–40 analyst-hours, plus data infrastructure.

Cost of a schema-filling pipeline run: under 10,000 tokens, fractions of a cent.

Share of the delivered artifact's substantive fields containing verifiable content: 0%.

Time I needed to establish that: four minutes.

The leakage is not the token cost. The leakage is the allocator downstream who reads a document with a header, a rating, and a confidence score, and mistakes the presence of a format for the presence of a finding. Every transaction is a potential extraction point, and the extraction here is paid in misplaced conviction rather than basis points.

Here is the part that will annoy the people who agree with me.

The pipeline that produced the null artifact was, in one narrow sense, behaving correctly. It refused to fabricate. It emitted empty fields rather than invented ones. That is a higher standard than a large share of the research currently circulating in this market — pipelines that generate full, confident, well-formatted reports whose claims have no traceable origin. An empty array is legible failure. You can see it. A hallucinated claim set is illegible failure: it resembles diligence exactly, and it costs the reader weeks to discover that the citations resolve to nothing.

In 2024 I spent two days tracing a Solana-based platform's ownership to a BVI shell with no physical presence in any regulated jurisdiction. What made that work useful was not the accusation. It was the paper trail. A pipeline that invents a paper trail is more dangerous than one that admits it has none. Logic holds; incentives collapse — and the incentive to sound confident is always stronger than the incentive to be blank.

The failure of my artifact is that it stopped at honesty. It should have refused to ship.

Ask your research vendor one question this quarter: how many fields in your standard output are permitted to return null and still pass the gate?

If the answer is "any," you are not buying analysis. You are buying a document with the correct number of headers. And in a market where liquidity is thinner than the last cycle and counterparties are thinner still, the illusion breaks when the liquidity dries up — usually for the allocator holding a report that was structurally incapable of saying no.