The table borders were perfect. Nine dimensions, forty-one rows, every cell populated. Technical assessment. Token economics. Regulatory exposure. Team and governance. Ecosystem position. Risk matrix. And inside every one of those cells, the same three characters: N/A.
I have reviewed smart contracts where the comments were more honest than the code. This was different. A research pipeline had been pointed at a source document. The ingestion layer returned an empty object, field by field, and the generation layer — trained, presumably, to be helpful — produced something shaped exactly like a due-diligence report.
Forty-one N/A fields. Zero bytes of evidence. And, near the end, a section headed "Minimum Input Required to Resume Analysis."
That is the part worth sitting with. The system knew it had nothing. It said so, eleven times, in immaculate Markdown, with a risk matrix and a signal-tracking table and a disclaimer explaining that its own failure did not constitute a negative judgment on any project.
Which, technically, is true. It could not have judged anything. It had never been given anything to judge.
I started this work in 2017, auditing bridge contracts for a chain that most of this industry has already forgotten. I was the only woman in the room and the youngest person in it, and I spent three weeks doing a line-by-line review that surfaced three reentrancy vulnerabilities the senior engineers had walked past because they were moving fast and believed they were right. That was the lesson of my career in one artifact: competence is the only currency, and it is paid in evidence, not in fluency.
Research used to be bottlenecked by human attention. Somebody had to read the whitepaper, pull the contract, cluster the wallets, and then sit with the discomfort long enough to publish a conclusion that would embarrass them if it turned out wrong. The bottleneck was not intelligence. It was the willingness to be embarrassed.
You cannot train that into a model, because the training signal rewards fluency, not falsifiability. So we now have research as industrial output. Agents write it, agents read it, agents summarize the summaries, and somewhere in that chain a human with a position size nods at a table and calls it conviction. In a sideways market this accelerates, because chop is exactly when nobody can be proven wrong quickly. Flat prices are a permission structure for bad analysis.
Forensics first. Three failure modes, ordered by how much damage they actually do.
Silent null propagation. Every pipeline has an implicit contract about shapes. The second-stage analyzer in this case did the correct thing — it detected an empty first-stage output and refused to invent the missing fields. That refusal is the only genuine signal in the entire document. But the pipeline still emitted a deliverable. Schema-valid. Semantically empty. In Solidity terms, this is the difference between a call that returns zero and a call that reverts. Both look like success to a caller that only inspects the return code and never checks the invariant. I spent most of 2020 staring at MEV bots that made precisely this mistake, treating a zero-value fill as a profitable route because the transaction did not throw. Capital and data follow the same physics: they move along the path of least resistance, and the dam is whatever formatting convention lets emptiness pass inspection as output. Liquidity flows like water, but greed builds dams.
Format affinity. Language models are rewarded for completion. A refusal is a short answer. A nine-dimension report is a long answer. Length reads as effort, and effort reads as competence, and so the frame got filled — not with invented numbers, which is ordinary hallucination, but with the systematic notation of absent numbers, which is more insidious because it is defensible. Read the document closely and you find something remarkable: it flags its own risk level as "cannot be assessed," then observes that this in itself constitutes the highest-grade process warning. That is a genuinely good sentence. It is also, structurally, an AI grading its own homework and awarding itself a pass for honesty. Trust is not a feature, it is a failed audit — and this was a self-audit with no independent verifier, which by definition produces no trust at all, only the appearance of it.
The checklist as governance theater. The report closes with a prioritized table of required inputs — red for mandatory, orange for important, yellow for nice-to-have — plus a list of signals to monitor and trigger conditions for restarting the analysis. It is a grant proposal. It is a remediation roadmap. It is, functionally, a DAO proposal. And that is where the comedy stops being about tooling. Turnout on substantive governance proposals across major protocols routinely sits in the low single digits; the quorum is met by a handful of delegates, some of whom are voting on each other's behalf under a delegation agreement nobody has re-read since the last unlock. A proposal to fix a data pipeline is a proposal to change human behavior, and behavior does not change because a table says priority: high. The checklist will be read by the person who already knew, ignored by the person who caused the problem, and cited later as evidence that due process was followed.
In 2026 I helped prototype an autonomous agent that negotiated micro-payments for on-chain data access. The negotiation logic was the easy part. The hard part was settlement of ambiguity: what does an agent do when it receives nothing, and must decide whether nothing was the answer or a failure? We wrote the rule ourselves — revert, never return zero. Then we shipped it and discovered that other people's agents had not.
Here is the counter-intuitive point, and I will take the downvotes for it. The null report is the most honest artifact the crypto research industry produced this cycle. Its competitors — the ten-page deep dives with a bull case, a bear case, a valuation range, and a roadmap table — where do you think their N/A fields went? They did not disappear. They were populated with the project's own marketing, aggregated by an agent that had no more evidence than this one did, and then formatted with confidence. The difference between those two documents is not information. It is aggression.
Everyone worries about models fabricating things. That is the wrong worry. Hallucination is loud and frequently falsifiable — an invented number can be checked against a block explorer. The failure mode that actually costs money is a default value that looks like a finding: a TVL figure with no timestamp, an "audited" badge with no disclosed scope, a governance health score computed from a snapshot that excluded the wallets that mattered. Transparency reveals the cracks that opacity hides, but a default value is opacity wearing a lab coat, and it passes review precisely because nobody thinks to challenge a number that appears in a table.
The second contrarian point is quieter. Sometimes the null is the truth. The source document genuinely contained nothing analyzable — no thesis, no protocol, no position to interrogate. The semantic layer was not broken and the analyst was not lazy. The schema was broken, which is a smaller and cheaper problem than everyone will pretend, and the correct institutional response is a revert condition, not a longer report.
The market corrects what the mind refuses to see, but it cannot correct what the format refuses to express. Within eighteen months, agent-authored research will carry provenance attestations — input hashes, retrieval logs, an explicit revert state — not because anyone will trust them more, but because we need to be able to distinguish a data failure from a taste failure when the position goes wrong. Volatility is the price of admission to the future. The silent return is the fee nobody itemizes.
