At 09:14 Zurich time, an automated research pipeline returned a document with 47 fields. Every one read N/A. Title: empty. Source: empty. Info-point list: empty. Nine analytical dimensions โ technical, tokenomic, market, ecosystem, regulatory, team, risk, narrative, supply chain โ all stamped "insufficient data."
The downstream model had a choice. It could have filled those slots with plausible fundamentals, invented a TVL figure, assigned a "low" risk rating, and shipped. Instead it printed the emptiness and flagged the plumbing.
An empty field is the loudest data point in any pipeline. That is not a philosophical position. It is an operational one, and it separates a research desk from a fiction desk.
Since 2024, every fund with more than three analysts has run some version of a two-stage LLM pipeline. Stage one ingests a source โ a news piece, a governance post, a token's docs โ and extracts structured info points. Stage two consumes those points and emits a standardized report across fixed dimensions. Stage one is extraction. Stage two is reasoning. The handoff between them is a JSON blob.
That handoff is where money gets lost.
I watched this failure mode from the other side in 2026. When I broke the NeuroTrade story โ an AI trading protocol whose "organic" volume turned out to be AI agents looping trades against each other โ the giveaway was never the volume number. It was the metadata. The wallets clustered in a pattern no human cohort produces. Synthetic demand leaves synthetic fingerprints. An empty field is the same class of evidence. It tells you something a filled version would have buried.
So here are the mechanics, and why they matter more than any single token.
Stage-one extraction fails silently. An LLM extractor does not crash when a source is unparseable โ it returns an empty object. A malformed article, a PDF that lost its text layer, a transcript that never arrived: all of them produce { }, and { } is valid JSON. Nothing throws. Nothing pages an engineer at 3am. The pipeline logs "success."
Stage two inherits the void. This is where discipline splits.
Most production systems are tuned for output completeness. Product managers measure report-generation rates. A model returning "insufficient data" looks broken; a model returning a confident nine-dimension analysis looks like it works. That incentive gradient is a fabrication engine. You get rewarded for filling blanks, so you fill blanks.
The analyst that refused did three things. It separated "no data" from "low risk" โ not the same thing, and conflating them is the most expensive error in quantitative research. An unknown risk is unbounded; a low risk is bounded. Collapsing the first into the second is how a desk ends up long a token it never actually assessed. It refused to infer where inference required a source: you cannot run a Howey test on a token with no jurisdiction, no team, and no distribution schedule, and any securities opinion written from a blank field is a hallucination in a suit. And it published the gap โ surfacing the plumbing break as the finding. In a market where information asymmetry is the product, a report that says "my feed is broken" beats one that says "here are nine things I invented."
Zoom out. This is not one pipeline. It is the industry's research layer, and the failure is systematic. Funds screen thousands of tokens with these systems. Exchanges auto-generate listing memos. Media outlets โ I have audited several โ auto-draft coverage from scraped sources. Each has a stage-one extraction feeding a stage-two generator, and each has some percentage of stage-ones returning { }.
The number that matters is not the failure rate. It is the fabrication rate โ the share of empty extractions that still produce a filled output. That number is unknown to everyone running these systems, including the people who built them.
I can tell you where it is highest: anywhere the success metric is article count, report volume, or coverage breadth. Wherever output is the KPI, the empty field gets painted over. Data is the only map I trust โ and maps are only as good as the survey behind them.
Source quality matters just as much, and almost nobody instruments it. Tether has published reserve attestations for years โ quarterly snapshots from an accounting firm, no defined scope, no liability, and no independent audit of the full balance sheet. The market treats those attestations as data. They are inputs with a confidence interval nobody computes. In early 2024, dissecting custody clauses in a spot Bitcoin ETF prospectus here in Zurich, I watched a comma-level language change move institutional risk appetite while every aggregator missed it. Fine print is where the real signal lives โ and it is exactly what a pipeline flattens into a single field.
Feed a pipeline an attestation and it will output certainty the underlying document never contained. Stage one cannot tell an audit from a screenshot. Stage two does not ask.
That is the whole game. Data quality sits upstream of reasoning quality, and almost nobody measures upstream.
Now flip the incident around, because the fear is pointed at the wrong target.
Everyone worries about AI hallucinating. Hallucination is loud. It produces claims that can be checked, caught, and punished. The quiet failure is an empty pipeline shipping a confident report. No anomaly triggers. No red flag. The output looks exactly like a good output. A fund sizes a position on a nine-dimension analysis of a token whose source data never loaded. The report is on file. The process was followed. Nobody did anything wrong, which is precisely the problem.
I have seen this pattern wear other clothes. The pitch that rollups need a dedicated data-availability layer was asserted, not measured โ when I pulled actual throughput, the overwhelming majority of rollups were posting volumes so modest that a purpose-built DA layer was overkill by orders of magnitude; nine out of ten of them would be fine on cheaper calldata. Hype is a trap; data is the only map I trust. Same with the "liquidity fragmentation" narrative VCs have re-run for three cycles. Fragmentation is a routing problem, and aggregators solved it in 2020. It keeps getting re-sold because a manufactured problem is easier to fund than a real one.
Arbitrage opportunities don't survive bad inputs. Neither do good ones.
If you run any part of a research stack, add one metric this quarter: the null rate at your stage-one handoff. Not the error rate โ errors are visible. The null rate. Then measure how often a null becomes a filled report downstream. That ratio is your fabrication coefficient, and it is the only number in your pipeline that tells you the truth about all the others.
The most honest thing a model did this month was print 47 fields of N/A. Watch whether anyone copies it.