Last Tuesday a due-diligence package arrived with a four-row table. Every cell read "not provided." Below the table, three hundred words of confident prose explained that the analysis had been completed and that conclusions would follow. Nothing followed. The vendor had shipped a template with its variables unrendered, then labeled the output a report.
I have spent months reading smart contracts that lie to me. That document was worse. It was a report that had never been told anything, wearing the costume of one that knew everything.
This is not an isolated glitch anymore. It is a product category.
The Market That Sells Verdicts It Cannot Produce
Through 2026, a new layer of "AI research agents" moved into crypto due diligence. The architecture is uniform across vendors. A scraper pulls the article, the docs page, or the governance forum thread. A retriever selects chunks. A model summarizes. A scoring layer assigns a confidence. The output is formatted into a PDF with headers, a risk matrix, and a recommendation.
The economics explain the speed. Extraction is cheap. Summarization is cheaper. Rendering a PDF costs nothing. Producing an actual evidence chain โ reading the arithmetic, reconciling the on-chain transfers against the foundation wallet, tracing who signed the multisig โ costs time, and time is the only input that does not compress.
So the industry compressed everything except the appearance of rigor. In a bear market where LPs are bleeding 30-40% quarter over quarter, the demand for a clean verdict is enormous. Sophisticated capital does not want a paragraph of caveats. It wants a number it can put in a memo.
The vendors supply the number. The caveats are the part that gets dropped when the template renders.
Where the Pipeline Actually Breaks
I traced this failure mode four times in the past month across different vendors. The break points are consistent.
First, the extraction layer returns an empty string, and empty strings are valid inputs. HTML body extraction fails constantly โ paywalls, canvas-rendered pages, aggressive consent overlays, Cloudflare interstitials. The function returns "". Nothing downstream rejects it, because no schema in the pipeline declares required: non-empty. A null propagates silently through retrieval, chunking, and completion.
Second, template interpolation converts absence into content. Every rendering engine has a fallback. When the variable is undefined, the fallback prints "N/A," "pending," or "no material findings identified." That last phrase is the dangerous one. It is grammatically identical whether the analyst found nothing or never looked.
Third, the language model will happily analyze nothing. This is the part people misread. A model asked to produce analysis from an empty context does not stall โ it produces a fluent essay about the absence of information, complete with hedging language and a confident conclusion. Perplexity is not evidence. Fluency is a property of the generator, not of the ground truth.
Fourth, the QA layer scores the wrong object. I have seen reports stamped "confidence: high" where the confidence score was computed over the output text โ its internal consistency, its coverage of the prompt's requested headings โ and never over the input. The model was highly confident about what it wrote. It was never asked whether what it wrote corresponded to anything.
Code does not lie; people do. The pipeline here was honest. It emitted a null at step one and carried it faithfully to step four. The dishonesty was inserted by a human who added a "finalize" button and a cover page.
The Asymmetry Nobody Prices
The cost structure of a false verdict is not symmetric. A false "unsafe" verdict costs a deal and some relationship capital. A false "safe" verdict costs the principal.
When I audited the 0x v2 exchange protocol in 2018, I spent four months on the maker fee calculation logic because the arithmetic did not reconcile at the boundary. Seven GitHub issues later โ one of them an integer overflow that could have drained liquidity pools โ the core team delayed mainnet by two months. That finding existed only because someone read the multiplication. No summary would have surfaced it. A summary of that contract reads "well-structured, standard patterns, no critical issues." The vulnerability lived in the gap between what the code did and what the spec said it did.
That gap is exactly what a summarizer cannot see, because the summarizer only reads the spec.
In 2026 I audited an AI-agent platform using crypto payments for autonomous service execution. The contracts had no audit trail for model decision-making โ no record of which input produced which action. The liability question was unanswerable because the artifact did not exist. Same failure class, different layer: the system produced outcomes without producing evidence.
High confidence on thin data is a warning, not a welcome. When a risk report arrives with a clean matrix and a strong recommendation, the first question is not whether it is right. It is what the input looked like. Ask for the extracted text. Ask for the source line numbers. If the vendor cannot produce them, the output is a rendering, not a finding.
The Null Was Correct
Here is the contrarian read, and it is not comfortable for either side.
The empty pipeline output was the system working. A pipeline that reports "insufficient input" is doing exactly what an honest instrument should do โ refusing to render. The failure was not the null. The failure was the organizational pressure that made the null unshippable, which is why somebody wrapped it in prose and sent it to a client.
The bulls are half right. AI will compress research costs dramatically. But it compresses the cost of producing text, not the cost of producing evidence. Those were always two different products sold under one invoice, and the market is only now discovering the difference because the text got cheap enough to expose it.
Forensics does not grade on intent; it grades on artifacts. The artifact here is a table with four empty cells. That is the entire evidentiary record of a document that claimed to know something.
Takeaway
The next time a risk memo lands with a recommendation and no citations, do not audit the recommendation. Audit the input. Ask to see the extraction log, the chunk boundaries, and the raw source strings that generated each claim. Audit the promise, not the poster.
A protocol that cannot tell you which block its numbers came from is not a protocol. It is a brochure. The research layer of this industry has spent two years building brochures with better typography.
One question remains open: when the next collapse produces a post-mortem citing "third-party analysis," how many of those citations will resolve to a document whose input was never non-empty?