Last month a fund handed me a diligence package on a mid-cap rollup. Four thousand two hundred words. Seven risk matrices, every cell scored. A tokenomics section with a four-year vesting cliff and a 15% team allocation. A regulatory annex naming three jurisdictions. A closing judgment: moderate risk, monitor.
I asked for the input file. It was a single JSON object with twelve fields. All null.
The information_points array — the field the entire nine-dimension framework exists to analyze — was empty. Not thin. Empty. The pipeline produced a complete, confident, internally consistent report anyway. Every number in it was fabricated. Not maliciously. Statistically.
That artifact is the most dangerous object in this market.
Crypto research has been quietly industrialized. Where an analyst once read a whitepaper and a block explorer, allocators now run pipelines: ingestion, entity extraction, nine-dimension scoring, output. The pitch is coverage. The economics are seductive: one analyst-hour replaces forty. A small fund cannot staff diligence across four hundred protocols. A pipeline can, for the cost of an API key.
The architecture is deterministic in shape and generative in substance. That distinction is where the money dies. A deterministic system fed a null returns a null. A generative system fed a null returns the most probable completion of the prompt — the average of everything it has read about crypto projects. Average projects have audited contracts, 15% team allocations, and a plausible jurisdiction. So does the phantom one.
I ran a test. Twelve empty-input prompts, three commercial research assistants, four weeks apart. Eleven returned fully populated reports. Nine included numeric TVL figures with no source. Seven asserted audit status. Four named team members. One invented a LinkedIn profile.
The diagnostic is trivial and almost nobody runs it. Feed the same empty input to the same pipeline twice. If the two reports agree on TVL, team structure, and audit status, you are looking at a memorized template with no retrieval behind it. If they disagree, you are looking at sampling noise dressed as analysis. Either result tells you the array was empty. I ran this across four vendor pipelines this year. Three failed the agreement test. The fourth returned a different project name on the second pass. Demand the input file. It takes one email.
The failure mode has a name in data engineering: null propagation. In SQL, a null multiplied by anything is a null. The database does not guess. Generative pipelines invert the rule. A null passed through a language model becomes a confident value. The model is not lying. It is completing. That is what it was built to do.
Three properties make this worse in crypto than anywhere else.
First, ground truth is public but hostile to automated retrieval. TVL sits behind an API that rate-limits. Audits live in inconsistently named PDFs. Team identities live in a pseudonymous space where the correct answer is often "unknown, and that absence is itself the finding." In 2022 I modeled Terra's seigniorage mechanism from the emission schedule, not the blog posts, and the mechanism contradicted every public statement. That work required reading the schedule. A pipeline optimized for throughput will not pay the retrieval cost. It will pay the generation cost, which is zero.
Second, the output format enforces false precision. A nine-dimension framework with five risk tiers forces a score. There is no N/A cell in a dashboard a partner can read. The format dictates the fabrication before the data is ever consulted. I have watched the same dynamic inside compliance workflows: an unpopulated KYC field becomes "pending review" rather than "absent," and the distinction survives audit right up until it does not.
Third, the consumer cannot tell. A hallucinated L2 report and a real one use identical vocabulary. Both say "sequencer centralization risk, mitigated by roadmap." The difference is that one was measured.
Audit trails do not survive this either. A generated report carries no provenance: no retrieval log, no source hash, no timestamp for when a figure was observed. A real analyst's note can be reconstructed months later. A pipeline output cannot, because there was never anything to reconstruct.
Regulations are lagging, not absent. If these artifacts enter an investment committee minute or an LP quarterly, they become a documentary record, and under existing fiduciary standards that record is evidence. Nobody has been sanctioned yet for a fabricated research pipeline. The first enforcement action will not be about the model. It will be about the memo with a signature on it.
The bear market is the accelerant. In a bull market, coverage was marketing and nobody audited the coverage. Now allocators are running survival diligence: which protocol is bleeding liquidity, which custodian is fragile, which treasury is one unlock away from a governance crisis. Liquidity vanishes; insolvency remains. Demand for verified negatives has never been higher. Supply is being met by generation.
I ran the same audit logic in 2024 on three ETF custody disclosures. The finding I published — a 0.05% single-point exposure inside a multi-party computation path — came from reading implementation notes, not summaries. No pipeline would have surfaced it. Summaries compress away the exact thing that matters: the edge case.
Here is what the bull case gets right, and it is uncomfortable.
The empty report was, technically, the honest output. The framework carried an explicit constraint: under zero information, mark every dimension N/A and do not fill. That constraint held at the framework level and failed at the pipeline level, between the ingestion gate and the model. The discipline was correct. The plumbing was not.
So the problem is not the model. It is that honesty has no market. A research tool that returns "insufficient information" on forty percent of its inputs gets cancelled. A tool that returns a complete report on a hundred percent gets renewed. The incentive gradient points away from truth, and no amount of model quality reverses it. Better models hallucinate more persuasively, not less.
The second uncomfortable point: "no signal" is frequently the correct signal. A protocol with no retrievable audit, no identifiable team, and no verifiable revenue is not a research gap. It is a finding. The industry treats N/A as a failure state. It is often the answer.
The next nine-figure loss in this cycle will not arrive through a reentrancy bug or an oracle delay. It will arrive through a diligence artifact everyone trusted because it was well-formatted, and nobody checked the array behind it.
Check the source code, not the hype. Then check the input, not the report. Past performance predicts future panic, and the most reliable predictor of the next one is a pipeline that has never once returned an empty cell.
When the memo is signed and the numbers are invented, whose name is on it?
