Last month I stress-tested a research pipeline that four crypto funds now license. I fed it an empty string. Not a redacted PDF. Not a broken scrape. Nothing. The system returned HTTP 200, a nine-dimension analytical framework, a six-category risk matrix, and a verdict of "N/A โ unable to evaluate." Every field was populated. Every field was empty. Latency: 3.2 seconds. Cost: forty cents. Had I not written the null-handling logic myself, that packet would have been forwarded to an investment committee as completed due diligence.
That is the failure mode nobody is pricing this cycle. Not the model inventing a fake token. The pipeline confidently reporting that it has nothing to report, while every human downstream reads "no flags" as "clean." Hunting for the story that defines the next cycle means first admitting that a growing share of the stories we are being sold is generated on top of empty inputs.
The 2026 bull market has manufactured a new asset class: the AI research agent. The pitch is uniform across a dozen funded teams. Point an autonomous agent at a token, a whitepaper, a governance forum, and it returns an institutional-grade memo in under a minute. The comparison to a junior analyst is made explicitly in the decks, always with a cost chart โ forty cents a run versus a salary. The sector's marketing leans on "Verifiable AI Compute": proof-of-inference, decentralized GPU markets, attestations that a given model actually ran a given computation on Render or Fetch-style networks. Market caps in this vertical have re-rated faster than any infrastructure narrative since restaking, and the GPU networks raised on the back of a story about trust that nobody has audited end to end.
I have watched this movie before. In 2021 it was "on-chain analytics dashboards" that rendered beautiful charts from a single mislabeled wallet cluster. In 2024 it was "AI copilots for trading" that backtested on survivorship-biased data. The technology genuinely improves each cycle. The failure mode does not. Everyone benchmarks the output. Almost nobody benchmarks the input. That asymmetry is where the money will be lost. An agent that produces a polished report from a broken scrape is not a productivity tool โ it is a laundering machine for garbage. And unlike a human analyst, it never says "I could not access the source."
Start with the anatomy, because the bug is architectural, not linguistic. A modern research pipeline has four stages: fetch, parse, schema-map, analyze. Each stage has a success path that does not require content to exist. The fetcher returns a 200 for a paywalled page whose body is a JavaScript shell. It returns a 200 for an anti-bot challenge page that reads "enable cookies." It returns a 200 for a document that loaded in the wrong character encoding, so the parser sees a wall of mojibake and dutifully counts it as tokens. The parser then extracts zero usable paragraphs and reports "0 tokens parsed" as a valid outcome, because in the developer's test suite that case was never wired to an error.
By the time the analyzer model receives its prompt, the object is structurally valid and semantically hollow. The deserializer is where the lie is manufactured. This is a type-system problem masquerading as a prompt-engineering problem. When a JSON key is absent, the mapping layer fills a default. An empty list for "information points." "Unclassified" for domain. "Not evaluated" for confidence. The downstream model cannot distinguish "I was handed nothing" from "the subject genuinely has no information points." Those are opposite facts wearing the same clothes. A human analyst reading a blank intake form knows the form is blank; the form's emptiness is itself information. The model reads a well-formed object with zero-variance fields and does what it was trained to do: complete the pattern.
And here is the part that should worry allocators. The most disciplined of these systems โ the one I tested โ actually behaved correctly. It refused to fabricate and emitted explicit N/A markers across all nine dimensions, with a note that unknown risk is itself the highest grade of uncertainty. That is the good case, and it still produced a document that looked like a report. Most commercial agents do not refuse. They interpolate. Given a template and an empty subject, a completion-trained model reaches for its priors: "projects in this category typically allocate 15โ20% to the team," "comparable L2s run 2,000 TPS," "median unlock cliff is twelve months." Plausible numbers enter the report wearing the costume of data, and the reader has no way to see the seam where fact stopped and prior began.
I have audited three of these pipelines over the past eighteen months, and the tell is always the same. Ask the operator to show you a run where the source document failed to load. Nine times out of ten they cannot, because their evaluation harness never tests it. The benchmark measures "did the agent produce a coherent answer," not "did the agent correctly decline." An agent that outputs "insufficient data" scores zero on nearly every leaderboard, so the incentive gradient points, relentlessly, toward always producing something. This is not a model-safety problem in the abstract. It is a measurable bias baked into how the industry scores its own tools.
The pressure compounds at the product layer. A vendor whose agent correctly declines gets churned by users who mistake refusal for incompetence. The customer asks for a report, the tool says "I have nothing," and the customer concludes the tool is broken rather than that the source was empty. So even honest teams are pushed toward fabrication by the same market that claims to want rigor. The incentive is not to be right. It is to be responsive. Those are not the same target, and in a bull market the gap between them is where capital quietly leaks.
Human-in-the-loop does not save you here, and it is worth being blunt about why. The human reviews the output, not the input. The reviewer sees a clean risk matrix with no red cells and approves it, because the seam is invisible by construction. The control that would actually work โ a person checking that the source document loaded โ has been designed out of the workflow in the name of speed. You cannot put a human in the loop of a process that hides the loop's broken link from them.
That is the quiet regulatory exposure too. In my 2025 compliance work with startups in Singapore and Vancouver, the recurring ask was a disclosure template an institutional counterparty would accept. The lesson we kept relearning: a disclosure that cannot prove its inputs is not a disclosure โ it is a liability with a cover page. An AI-generated risk memo with no provenance chain is, from a diligence standpoint, worse than no memo at all. It creates the appearance of rigor while removing the human who would have noticed the void. A regulator will eventually ask a simple question โ show me the document this analysis was built on โ and "the model interpolated from priors" is not an answer that survives a hearing.
Now quantify the decision pollution, because this is where the abstraction becomes a P&L line. Consider a fund running forty positions at 2% each. Suppose three of the forty due-diligence packets were generated from zero-signal inputs but read as clean โ no red flags in the risk matrix, because "N/A" was rendered in the same green as "low." That is 6% of the book allocated on the strength of a document that contains no information. In a market where a single mispriced position can define the year, a 7.5% silent-failure rate in the research layer is not a rounding error. It is the whole game, played blind.
The semantic crime is specific and worth naming. "N/A" is not "low risk." It is the highest-risk state in the entire matrix. Not-evaluated is not the absence of danger; it is the absence of knowledge, which is a strictly worse position. The framework that produced my test report even said so, in bold โ unknown risk is the highest grade of uncertainty, do not misread it as a safety signal. Nobody reads that line. They read the matrix, see no red cells, and move on. The pipeline optimized for a clean dashboard, and a clean dashboard is exactly what an empty input produces.
So what breaks the loop? Not better prompting. Better plumbing. The missing primitive is verifiable input provenance โ the mirror image of the verifiable-inference narrative everyone is chasing. Content-addressed source documents. Signed fetch receipts that fail closed. A hard input-validation gate that returns an error code, not an analysis, when the information-point list is empty. In software terms this is trivial: assert that the information-point list is non-empty before the analyzer is ever invoked. In market terms it is the difference between a research tool and a random-number generator with good typography.
The teams that internalize this will build a moat the hype cycle cannot copy, because the moat is not a model. It is the discipline to say "I was given nothing, and I will not pretend otherwise." That discipline is unglamorous, it does not demo well, and it is the only thing standing between an institution and a six-figure allocation made on forty cents of compute and zero bytes of signal.
Here is the blind spot the whole sector shares. Everyone is racing to make inference verifiable โ proving that the model ran, that the GPU was honest, that the weights were intact. Admirable, and largely beside the point. The expensive failure is not unverifiable output; it is unverifiable input. You can have a perfectly attested computation running on a cryptographically pristine GPU over a document that never loaded, and the attestation will certify the garbage with the same confidence it would certify the truth. Proof-of-inference secures the wrong end of the pipe. We are building notarized seals for the answer while leaving the question unexamined.
The contrarian corollary is uncomfortable for the product people: an agent that refuses to answer is a feature, and the market is pricing it as a bug. Every dashboard that rewards "always produces output" is training its users to trust silence. The next generation of serious research tools will be judged not by how much they generate but by how cleanly they fail. Watch for the first team that puts "correctly declined" on its benchmark and charges a premium for it. That is the tell that the adults have arrived. Hunting for the story that defines the next cycle, I would rather back the tool that tells me it has nothing than the one that tells me everything.
The next narrative is not agents that analyze. It is agents that can prove what they were given โ provenance before intelligence. When that primitive ships, the zero-signal report becomes impossible to fake, and half the current sector's market cap re-rates against the other half. The question every allocator should ask their research vendor this quarter is not "how good is your model." It is: show me the last time it said no.