Empty Input, Zero Hallucination: What a Failed Analysis Pipeline Reveals About AI-Driven Due Diligence

The output was a fully formatted nine-dimension analysis report. Every table was rendered. Every section carried a header. And every single field read the same three letters: N/A. That is not a bug in the report — it is the report working exactly as designed, and it is the most honest artifact I have reviewed all quarter. Here is what happened. A two-stage research pipeline was constructed to evaluate a crypto asset. Stage one was supposed to extract structured information points from a news article: project name, technical architecture, token supply, market data, team background. Stage two was then supposed to analyze those points across nine dimensions — technical feasibility, tokenomics, market positioning, ecosystem role, regulatory exposure, team and governance, risk, narrative, and supply-chain contagion. Standard machinery. I have built versions of this myself for audit triage. Stage one returned an unfilled template. Every field was marked 'not provided / unclassified / not evaluated.' The information-point list was empty. No title, no source, no body text. The pipeline's second stage then did something that most production systems fail to do: it refused to proceed. It emitted a complete analytical frame with every dimension tagged N/A, and appended — for each dimension — the specific inputs required to activate it. Contract address and audit report for technical review. Tokenomics chart and unlock schedule for supply analysis. Legal entity and registration jurisdiction for the Howey test. The report closed with a status line: BLOCKED, input empty. Contrary to the prevailing assumption that large-model pipelines degrade into confident nonsense under thin inputs, the failure mode here was the opposite. The system produced zero fabricated claims across roughly 1,400 words of structured output. The cost of that discipline is obvious: no information value, a five-star rating of zero across the board. The benefit is subtler and, for anyone deploying these tools in capital allocation, more important. Consider the alternative. Given an empty extraction, a less constrained model will reach for priors. It knows the typical shape of a DeFi token launch — 15% team allocation, four-year vesting, a 40% initial float. It knows the standard rollup architecture — sequencer, prover, data availability layer. These priors are statistically reasonable and individually plausible. Stitched together into a due-diligence memo, they become a hallucinated asset profile that reads exactly like a real one. I have seen this failure up close. In my 2017 Solidity audit work at a São Paulo fintech, the dangerous code was never the obviously broken function; it was the function that looked correct because it followed the template. A fabricated supply schedule is the analytical equivalent of a reentrancy bug that passes a surface-level read: structurally convincing, catastrophically wrong at the point of execution. The economic logic maps cleanly. An analyst who writes 'N/A — insufficient input' pays a visible cost in perceived productivity. An analyst who writes a confident memo built on inferred priors collects the same fee with none of the friction. Multiply that incentive across every AI research tool now being sold to funds, and you get a market that rewards the appearance of coverage over the reality of it. The empty-input report is a rare instance of a system pricing its own ignorance correctly. But the contrarian read cuts against the pipeline's builders, not just its imitators. A refusal that produces 1,400 words of N/A scaffolding is safe, and it is also nearly useless. The genuinely hard engineering problem is not teaching a model to say 'insufficient data.' It is teaching the model to distinguish between data that is absent and data that is merely unstructured — buried in prose, implied by a funding announcement, recoverable from an on-chain query. The report's activation conditions are honest but lazy: 'needs contract address' is a request for human labor, not a research capability. A better system would have flagged that stage one itself had malfunctioned — that a template returning uniformly empty fields is a pipeline fault, not a source-document fault. The report noted this possibility in its risk section and then did nothing with it. Logic is binary; intent is often ambiguous, and here the ambiguity sat unexamined in the diagnostic layer. That distinction matters more as these tools move from summarization into allocation. When an AI agent is asked whether to size a position, 'I have no data' and 'I have data I failed to parse' demand completely different responses: abort versus debug. A system that collapses both into the same N/A output is safe but blind. The next generation of due-diligence infrastructure will be judged less on whether it avoids hallucination — that is becoming table stakes — and more on whether it can tell an empty well from a broken pump. The report in front of me passed the first test and never attempted the second. That is progress. It is not yet analysis.