Eleven crypto research reports crossed my desk last quarter, produced by four separate AI analyst pipelines: two at funds, one at a DAO treasury committee, one subscription desk. Four of the eleven contained at least one section where every field resolved to the same string โ N/A, insufficient information. All four shipped anyway, stamped "analysis complete," and were forwarded to people with signing authority over real capital. The error rate was not the story. The fact that nobody downstream caught it was.
I pulled the full diagnostic from one of them. It is a Stage Two Deep Analysis โ the nine-dimension teardown now standard in crypto research: technology, token economics, market structure, ecosystem position, regulatory exposure, team and governance, risk, narrative, supply-chain transmission. The document opens with an input-integrity gate. The gate failed. Title missing. Source missing. Article type unclassified. Information point list empty โ an empty array, the one field the entire analysis chain hangs from. Then the pipeline did exactly what it was instructed to do: it rendered all nine dimensions, filled them with N/A, and delivered.
The marginal cost of a report went to zero. The marginal cost of being wrong did not.
In 2023 a nine-dimension report cost an analyst two days and roughly $1,200 in billable time at a mid-tier fund. In 2026 the marginal cost of the same document is under one cent. Report volume at the desks I track is up roughly 40x since 2023 while research headcount is flat to down. Volume is what gets reported upward, because volume is what looks like coverage. There is no line item on any dashboard for "the report had nothing in it."
The schema itself is not the problem. It mirrors what a competent human analyst does. The problem is that the schema specifies what to output, not what must be true before outputting. Constraint seven in the pipeline brief said: output all nine dimensions in full. The model obeyed constraint seven and violated the intent of the whole document. That is Goodhart's law running in production at forty times the volume, and it is now the dominant failure mode in on-chain research. Not hallucination. Compliance.
The distinction the pipeline could not make
N/A because nothing was found, versus N/A because nothing is there. To a reader those render identically. To a capital allocator they are opposites. The first means unverified โ do not size a position. The second means checked, clean.
In the document I pulled, token economics returned N/A on supply schedule, unlock cliffs, and incentive sustainability. Team and governance returned N/A on everything. Those are precisely the two dimensions that require primary data โ a token contract, a vesting schedule, a multisig signer list. N/A in those two slots is a fingerprint. It means the pipeline never ingested the source at all, and every downstream number inherited that void.
To its credit, the document diagnosed itself. It isolated the exact break point in the evidence chain and flagged a meta-risk: the danger of downstream systems defaulting empty fields to "no risk found." The analysis was technically correct. It simply had no authority to stop anything from shipping.
I have seen this bug in production contracts, and the parallel is structural, not decorative. An oracle that returns zero instead of reverting does not break the calling contract โ it completes it, with garbage. A lending market that reads a zero price does not throw; it liquidates the entire pool. Chainlink's consumer design exists because of this: latestRoundData returns an updatedAt timestamp specifically so the consumer is forced to check staleness, because the producer will never validate itself. Six years of oracle exploits taught the same lesson. Validation cannot live upstream.
Research pipelines have the same shape. The producer โ model, prompt, schema โ has no mechanism to refuse delivery. "Input integrity failed" is not one of the nine dimensions, so failure has nowhere to be reported. The gate can only live at the consumer: a hard schema constraint, not a prompt instruction, requiring a minimum of three numbered information points before analysis is permitted to run. Below that threshold the pipeline returns an explicit VOID and routes back to source collection. Not a report with placeholders. A logged, counted refusal.
Everyone will blame the model. The model behaved correctly.
The fix vendors will sell is better prompting โ instruct the model to refuse when input is empty. That is not a fix. Prompt-level instructions are advisory; schema-level constraints are enforceable. The gap between telling a system not to proceed and making it structurally unable to proceed is the gap between a require statement and a code comment reading "do not transfer more than balance."
I spent 2017 auditing ERC-20 contracts by hand at a Singapore fund โ fifty-odd of them before the ICO wave crested. Three had reentrancy vulnerabilities severe enough to reject the position, and rejecting those three saved the firm roughly $2 million when the market broke. The lesson that stuck was not about reentrancy. It was that the contracts that actually cost people money were rarely the ones with obvious flaws โ those got caught, because someone was looking. The expensive ones had clean checklists. The auditor reached the bottom of the template and stopped, because the template was full.
That is the blind spot here. The model declined to fabricate. It is the only component in the stack that did its job. The human layer read a full template and concluded the work was done. N/A got parsed as neutral โ "nothing alarming found" rather than "nothing found at all" โ because a finished document produces the same psychological closure as a finished audit.
Underneath sits an incentive problem. Research desks are measured on coverage and cadence, not evidentiary density. A desk shipping 200 hollow notes outranks a desk shipping 40 verified ones on every dashboard that counts output. Nobody is paid to publish the void.
What to do with this
When a research note cites a source document that itself flagged missing input, the report is not partially valid โ it is void. Not discounted. Void. Two checks separate real analysis from a filled-in template: can every conclusion be traced to a numbered information point, and does the token economics section contain an actual number rather than a dash. If a report shows N/A in token economics or governance, assume the pipeline never read the source, then ask why it shipped.
Smart money doesn't read the conclusion first. It reads the sourcing.
The next cycle will not be decided by which desk produces the most reports. It will be decided by which desk can trace a claim back to a block height โ and can prove it didn't skip the gate to get there.