The report landed at 6:40 a.m. Nairobi time. Nine sections, formatted to spec. Technical analysis. Token economics. Market structure. Ecosystem position. Regulatory exposure. Team and governance. Risk matrix. Narrative. Supply-chain transmission.
Every single one of them stamped N/A - information insufficient.
Not "data pending." Not "sources unconfirmed." Forty-seven separate instances of the same placeholder string, arranged into tables so tidy they looked deliberate. The pipeline had not crashed. It had not timed out. It had shipped.
I read it twice before I understood what I was holding. This was not a broken product. It was a product that had been asked to analyze nothing, and had answered honestly, in nine dimensions, at length, without ever raising an error.
Here is the plumbing. Two-stage LLM pipeline โ the kind every fund, newsroom and research desk in this market now runs. Stage one ingests an article and deconstructs it into a structured array of information points: facts, sources, project names, time sensitivity. Stage two takes that array and runs it through a nine-dimension analytical framework, outputting technical, tokenomic, market, ecosystem, regulatory, governance, risk, narrative and transmission assessments.
Stage one returned an empty array. No facts. No sources. No project names.
Stage two ran anyway. It built the full framework, populated every cell with the same three-word refusal, and appended a remediation checklist telling the operator precisely which seven fields to resubmit โ six marked "recommended," one marked mandatory.
The whole document is a masterclass in a failure mode that the AI-agent narrative of 2026 has no vocabulary for. We spent two years arguing about hallucination, about models inventing facts. Nobody built the slides for the opposite problem: models that follow the template so faithfully they manufacture the shape of analysis with none of the substance. The silence after the funding round tells the real story, and this silence arrived in a table with a header row.
And this is happening in the middle of the largest capital rotation this industry has ever seen. I sat in a Nairobi roundtable last quarter between local fintech founders and European regulators, and half the due-diligence decks on the table were machine-assembled. Institutional money entering African markets is not reading whitepapers anymore. It is reading pipeline output.
I want to be precise about what went wrong, because the obvious diagnosis is wrong.
The model did not hallucinate. It did the opposite. Handed an empty array, it refused โ politely, systematically, in nine columns โ to invent a single technical claim. No fabricated TVL. No invented unlock schedule. No imaginary validator count. Under a strict reading, this is the most well-behaved agent I have encountered all year.
The failure sits one layer up, in the schema. Two specific defects.
First: there is no input gate. In any multi-stage LLM system, the standard engineering practice is a validation node that halts the chain when the upstream payload is empty or malformed. The check is trivial โ if the information-point array has zero length, break and alert. It costs nothing. It was absent. So a null payload flowed downstream, and the downstream stage, which is optimized to complete its task, completed its task.
Second: a circular dependency in the field specification. The "projects and protocols involved" field was defined as something to be identified from the information-points list. With the list empty, the field had no upstream source and no independent derivation path. That is not a model bug. That is a spec bug โ a logical deadlock written into the requirements document, invisible until the day the list went to zero.
I have made both of these mistakes myself. In 2021, at a generative art viewing in Mombasa, I praised a project's roadmap on the strength of a casual conversation and published before I read the contract. It was a honeypot. The backlash cost me real credibility, and I rebuilt it on a livestream where I walked through my own error line by line. Two-source verification has been mandatory on every scoop since.
What I learned that week was that my failure mode was enthusiasm. I filled silence with narrative. The pipeline in front of me did not do that. It left the silence in place.
The test I now run on any agent stack takes about four minutes. Feed it a deliberately empty payload and watch what comes back. A system with an input gate stops. A system without one returns a beautifully formatted document describing nothing. Then feed it a payload containing one plausible but unverifiable claim, and see whether the output flags the claim or absorbs it into the risk matrix as established fact. The first test catches the null case. The second catches the case that will actually cost you money.
Now the part that should worry you more than any individual bug: why nobody caught this. Pipelines are measured on throughput and uptime. A stage that halts on bad input registers as a crash on the dashboard. A stage that emits nine sections of N/A registers as a successful run โ 200 OK, four thousand tokens, zero exceptions. The system that fails silently looks healthier in the metrics than the system that fails loudly. That is not a bug in the pipeline. That is a bug in how we instrument it, and it is industry-wide.
There is one detail in the remediation table I keep returning to. The report listed seven fields required to re-trigger analysis. Six were "recommended." One was mandatory, and it was the information-points list itself โ at least three to five structured entries, each with a source. The failing system had correctly identified the exact minimum viable input it needed, and had asked for it in writing.
A pipeline that documents its own failure conditions is a pipeline someone built with care. The missing gate wasn't carelessness at the model layer. It was a gap between two teams โ the people who wrote the spec and the people who wired the stages โ and the gap had a shape, and the shape had nine columns.
Everyone who hears this story will draw the comfortable conclusion: AI research tools aren't ready, the agents are unreliable, wait for the next model. Wrong read.
The dangerous reports are not the ones that came back empty. They are the several hundred that came back full. Same thin input, same two-stage architecture, same missing gate โ except those runs happened to scrape a paragraph dense enough to feed the model something, so the framework filled itself with plausible, sourced-looking, unverified prose, and nobody flagged anything because nothing looked broken.
We only caught this one because the input was so catastrophically empty that even a compliant model had nothing left to launder. That is a random audit, not a control. The silence after the pump tells the real story, and here the silence was loud enough to survive formatting.
The commercial incentive makes it worse. If my agent returns N/A and yours returns "strong buy, accumulate on dips," which one gets renewed? Which one gets quoted on a panel? Placeholder text becomes analysis the moment it is pasted into a slide deck, and it becomes consensus the moment two decks cite each other. Nobody in this cycle is being paid to refuse.
So here is what I am watching. Refusal rate, as a due-diligence metric. Ask any vendor, any research desk, any fund running agents: what percentage of your pipeline runs halt on insufficient input? If the answer is zero, they do not have an input gate. They have a word generator with a schema bolted on top.
The silence after the demo tells the real story. And right now, almost nobody is measuring it.