Contrary to popular belief, the most alarming document to cross my desk this quarter contained no false claims. It contained no claims at all.
A two-stage due-diligence pipeline โ the kind now marketed to funds as an 'autonomous research analyst' โ returned a 3,000-word report in which every material field read N/A โ insufficient information. Technical maturity: N/A. Token unlock schedule: N/A. Howey test factors: N/A. Team assessment: N/A. A complete, formally rigorous, nine-dimension forensic framework, applied to an empty input.
The document was not a failure of analysis. It was a failure of infrastructure, dressed as humility. In a market where capital moves faster than verification, that is the failure mode nobody is pricing.
I have spent nineteen years watching research frameworks promise rigor and deliver theater. This one promised theater and delivered โ by accident โ the single most honest artifact I have read this cycle. The reasons why are worth dissecting.
Context: the industrialization of verification
The due-diligence market has changed shape. Three years ago, a fund's research process was a senior analyst, a Bloomberg terminal, and a Notion page. Today it is a pipeline: an ingestion stage that scrapes whitepapers, governance forums, and on-chain data; an extraction stage that converts prose into structured 'information points'; and an analysis stage that applies a fixed framework โ tokenomics, ecosystem position, regulatory exposure โ to each point.
The economics are obvious. A junior analyst costs $120,000 a year and can cover perhaps fifteen projects to depth. A pipeline costs a fraction of that and claims to cover three hundred. Under bull-market conditions, where a single mispriced launch can return 40x, the temptation to buy coverage breadth instead of coverage quality is overwhelming.
I built a version of this stack myself in 2020, simulating the Curve 3Pool under a 15% depeg. The simulation was the easy part. The hard part was the ingestion layer โ the unglamorous plumbing that turns unstructured text into fields a model can reason over. That layer is where every pipeline I have audited quietly fails.
The bull market accelerates the pathology. In euphoria, funds do not buy research to find reasons to say no โ they buy it to document reasons to say yes, faster. A pipeline that returns a formatted report in ninety seconds satisfies that demand regardless of whether the report contains a signal. Speed has been decoupled from verification, and the coupling has not been restored.
Core: what the N/A report actually exposes
Here is the mechanical failure, reconstructed.
The pipeline is two-stage by design. Stage one โ extraction โ is responsible for parsing source text into structured information points: project name, technical claims, token distribution, involved parties. Stage two โ analysis โ consumes those points and emits a nine-dimension judgment. The architecture is sound in principle. Separation of concerns. Extraction should be testable independently of interpretation.
But the two stages were wired together with no contract. No schema validation. No assertion that the fields stage two consumes were actually populated by stage one. When stage one returned an empty information-point list โ whether from a parsing failure, a transmission error, or a genuinely empty source document โ stage two did not halt. It proceeded. It walked its framework end to end, filling every slot with a formal placeholder: N/A โ information insufficient.
This is the critical defect. A research pipeline that cannot distinguish 'no data' from 'negative data' has no analytical integrity, only formatting integrity. Stage two treated absence of evidence as evidence, then wrapped the void in the visual language of institutional rigor โ tables, star ratings, risk matrices, Howey-test checklists โ all rendered with the confidence of a completed audit.
Follow the artifact. An institutional reader receives it. The report is formatted. It has sections. It cites a framework. The eye, trained to scan for red flags, finds compliance. What it does not find โ because nothing in the layout signals it โ is that the entire document is an empty set.
This is not a hypothetical. I have seen the downstream decision. In late 2024, reviewing custody specifications for the spot Bitcoin ETFs, I encountered a diligence memo where three of seven issuers had multisig implementations that differed from their public descriptions. The memo flagged none of them. The researcher had relied on an upstream summary that had, in turn, relied on a pipeline that had, in turn, failed silently on the custody section. Ownership is an illusion without immutable proof โ and so is a risk assessment without a verified input chain.
The tokenomics table in the N/A report is the cleanest illustration. It has four rows: team, early investors, community and liquidity, treasury. Each row carries three columns: allocation, unlock schedule, risk flag. Every cell reads N/A. A reader skimming for the unlock schedule finds the row, finds the column, finds a value โ N/A โ and files it as 'unknown,' which the brain processes as 'to be determined later.' The report has laundered a catastrophic data failure into a neutral placeholder.
The regulatory section is worse. A Howey test rendered as four rows โ money investment, common enterprise, expectation of profit, efforts of others โ each marked N/A, with a composite verdict of 'unable to assess.' The framework's own honesty becomes the camouflage. Because the pipeline is explicit about its ignorance, it reads as cautious rather than broken.
I stress-tested the failure myself. I built a minimal two-stage replica in Python โ a parser returning an information-point list, an analyzer consuming it, and deliberately no assertion in between. I fed the parser fifty whitepapers, then severed the connection and fed the analyzer nothing. It produced fifty identical, fully formatted reports. Every one passed a naive schema check, because the schema validated shape, not population. A report full of N/A strings satisfies a regex that expects strings.
The compliance layer launders the same way. Most project KYC is theater: a tiered verification gate that buys the appearance of diligence while the actual capital โ wallet holdings โ crosses the boundary untouched. Compliance costs are passed to honest users, who submit passports, while a few concentrated wallets bypass the gate entirely. The N/A report is the same structure. It is verification theater for research: the appearance of a diligence process, at zero verification cost, with the burden shifted to the reader who assumes the fields were populated.
There is a specific asymmetry here that funds are not modeling. A confabulated report fails loudly: the numbers do not match reality, and someone eventually notices. An N/A report fails silently: it looks like a completed process, it triggers no alarm, and it is filed. Silent failure is more expensive than loud failure because it defers the cost to the decision it never informed.
Consider the composite verdict. The framework's nine dimensions terminate in a core judgment and an information value rating โ stars, from one to five. The N/A report awarded itself one star per dimension: technical value, investment value, timeliness, reference. Four one-star ratings, presented with the neutrality of an assessment. The framework assigned the void a score. A score is an opinion. The pipeline emitted an opinion about nothing and formatted it as a finding.
Now the accountability question. A human analyst who submitted this report would be fired. The human, after all, had a source document and one job. The pipeline has no such accountability layer. It has no owner, no signatory, no line where a name attaches to a claim. When it fails, the failure is distributed across an ingestion service, a parsing model, a prompt template, and a rendering function โ none of which will be held to anything.
Institutional custodial skepticism exists because custody concentrates liability. The pipeline does the opposite: it disperses liability until it evaporates.
Zoom out to the industry level. The transmission chain matters more than any single report. Pipeline output flows to an analyst, who condenses it into a memo; the memo flows to an investment committee, which approves a position; the position enters a fund's net asset value. Each handoff strips context. By the time the N/A reached the committee, it was a one-line bullet: 'no red flags identified.' The absence of a red flag became a green flag. The interface between stages โ human and machine alike โ is where integrity leaks, not within any single stage.
This is why I keep returning to primary sources. A whitepaper, a contract, a governance thread โ these are immutable in a way that summaries are not. Every layer of abstraction added between the source and the decision increases the surface area for silent failure. The pipeline added three layers. The N/A report had four. Ownership is an illusion without immutable proof โ and so is diligence without a verifiable chain back to the raw artifact.
Contrarian: what the pipeline got right
Here is what the bulls โ and the pipeline's defenders โ get right, and it deserves weight. The system refused to hallucinate.
Every commercial LLM I have stress-tested under these conditions will, when handed sparse input, fabricate. Given a project name and nothing else, it will invent a token distribution, a plausible team, a roadmap. It will fill the void because the void is an error state and models are trained to avoid error states. I have run this test dozens of times. The confabulation rate on empty inputs exceeds 80%.
This pipeline scored zero. It output N/A nine times and refused to improvise. That is not a bug. That is the rarest behavior in AI-generated research: the discipline to say 'I do not know' when the penalty for honesty is an unusable output.
Most failures in this market are the opposite โ full reports, confident prose, invented numbers that a fund deploys against. An empty report wastes a meeting. A confabulated report wastes capital. Scored honestly, the N/A pipeline is the safer failure.
The defenders are also right that the framework itself is sound. Nine dimensions, correctly populated, constitute genuine diligence. The starvation was upstream. Fix the ingestion contract โ assert that stage one produced non-empty fields before stage two runs โ and the architecture holds.
Takeaway
The dangerous artifact is not the pipeline that says N/A. It is the pipeline that says 'here is the analysis' when there was nothing to analyze โ and the analyst who signs it anyway. As coverage scales toward every launch, every fork, every governance proposal, the question is not whether your research can be automated. It is who signs the output, and what schema prevents a void from being rendered as a verdict. Verify the input before you trust the report. The report cannot verify itself.
