Last month my desk ran a routine diligence pass on a mid-cap DeFi protocol. Standard four-stage pipeline: fetch, extract, analyze, render. The fetch returned HTTP 200. The extractor returned an empty array. The analyzer returned nine full dimensions of analysis โ technical, tokenomics, market, ecosystem position, regulatory, team and governance, risk, narrative, supply-chain transmission โ wrapped in forty-one tables, a six-row risk matrix, and a Howey test breakdown.
Every substantive cell said N/A.
Nobody on the call flagged it. Two analysts read the document end to end and asked a follow-up question about the regulatory exposure section. There was no regulatory exposure section. There was a table header, and beneath it the string "N/A - insufficient information," repeated eleven times, rendered in a font that looked expensive.
That is the most expensive bug in crypto research right now. Not hallucination. The container.
The architecture nobody audits
Crypto research has been re-platformed onto LLM pipelines in roughly eighteen months, and the architecture is now depressingly uniform. Stage one fetches the source. Stage two decomposes it into atomic, individually verifiable claims โ what the pipeline designers call information points. Stage three runs those points through analytical dimensions. Stage four renders.
Alpha is captured or destroyed at stage two. Analysts spend all their time reading stage four.
I watched this happen from the inside during the Terra collapse. My team pulled on-chain data from twelve major wallets and mapped the exit sequence days before the public narrative cracked. The edge was not the analysis. The analysis was arithmetic once the extraction worked. The edge was that we had actual transaction hashes in hand while everyone else had a Twitter thread. Never trust the narrative, only trust the wallet history โ and never trust a report that has no wallet history in it.
Here is the diagnostic that took me two minutes and should have taken the room thirty seconds: the null-input report contained zero proper nouns. Not one contract address. Not one ticker. Not one block height, one funding round, one jurisdiction, one named contributor. A nine-dimension analysis of nothing, in a document with no nouns.
That is the tell. A report with no verifiable identifiers is not a low-confidence report. It is not a report.
Four ways the fetch lies to you
HTTP 200 means the server responded. It does not mean you received the article. I have now catalogued the four failure modes that produce a silent empty extraction, and every one of them returns success at the transport layer.
Paywall intercept. The request returns 200 with a subscription page. Length looks plausible. Extraction yields marketing copy about the value of journalism.
Image-only document. The PDF is a scan. There is no OCR layer. Extraction yields whitespace and, if you are unlucky, the footer.
Encoding mismatch. The body arrives in an unexpected charset. The extractor discards the undecodable majority rather than failing, because failing would break the pipeline.
Length truncation. The tail is dropped at a token boundary. You get the first third of a research report and never learn the conclusion contradicted the abstract.
None of these raise an exception. All four are indistinguishable from success until a human reads the output and notices the nouns are missing. On a desk processing forty documents a day, that human check is the first thing to go.
The instruction that caused it
The pipeline did not hallucinate. Read the artifact carefully and the model's behavior was impeccable. It was instructed to preserve the analytical framework and populate unavailable fields with an explicit insufficiency marker. It complied, forty-one times, with the correct string every single time.
The bug is the instruction. Somebody, reasonably, wanted a stable output schema so downstream systems could parse it. So the schema was made mandatory and the content was made optional. That inversion is the whole failure.
A table with a header reads as analysis to a human skimming at two hundred words per second. The visual grammar of research โ column rules, bold verdicts, confidence tiers โ is doing the persuading. The cells are empty. Structure is a credibility multiplier, and we handed it to a system with nothing to say.
Here is what my team runs instead. Not a suggestion. A gate.