On a Tuesday morning in Toronto, a research pipeline I built returned a nine-section report. Every cell read the same line: N/A โ insufficient information. No technical assessment. No token supply table. No risk matrix. Just a clean, complete, empty skeleton, and one sentence at the bottom: input validation failed.
I read it twice. Then I did the thing I almost never do with machine output โ I approved it.
Three weeks earlier, a sibling pipeline had been fed a similarly broken feed. Stage one extracted nothing. No title. No source. An empty array where the information atoms should have been. Stage two, asked to analyze that emptiness, produced 1,800 words of confident prose. It invented a supply schedule. It rated the governance risk of a model that did not exist. It was forwarded to a desk, and it shaped a position.
Same code. Two runs. One lie, one refusal. The refusal was the more valuable output, and nobody wanted to read it.
The bull market has industrialized research. Every fund, every desk, every alpha channel now runs some flavor of automated extraction โ scrape, decompose, score, publish. The economics are obvious. A human analyst costs a salary and produces two deep dives a week. An agentic pipeline costs API calls and produces two hundred. Volume became the product. Nobody audits the input contract anymore, because the input contract is boring, and boring does not raise a fund. Volume is not information gain. It is close to the opposite โ the more briefs a desk produces, the less any single one of them gets read.
Meanwhile the substrate got faster. I spent much of the past year on a prototype where AI-driven agents settled micro-transactions on a modular chain, batching calls to cut gas by roughly 40%. The lesson was not that agents are efficient. The lesson was that agent efficiency is entirely contingent on the execution layer being trustless โ and completely indifferent to whether the data being reasoned over is true. A bot that settles in 400 milliseconds on fabricated inputs is not a fast bot. It is a fast error.
The bull market makes this worse in a specific, mechanical way. Funding rounds no longer require working code. I have watched a project raise nine figures this quarter on a testnet that has never processed a transfer above the dust threshold, and the research brief accompanying that raise carried no fewer than fourteen unsourced market-size figures. Every one of them traced back to the same three vendor reports, which traced back to each other.
Settlement latency has collapsed. Verification latency has not.
Strip an automated research pipeline to its bones and you find three layers, and only one of them fails loudly. The architecture of trust, stripped to its bones, is really just a sequence of format conversions.
Layer one: extraction. Consumes a source, emits structured information points. Layer two: analysis. Consumes those points, emits judgment. Layer three: publication.
The failure is structural: an empty array is syntactically valid. A schema demanding information_points: array is satisfied by []. The type check passes. The pipeline runs. Nothing throws. Stage two receives a well-formed object containing zero facts and is then instructed by its own prompt to "produce a complete eight-dimension analysis."
Of course it hallucinates. You have handed a completion engine a mandatory output format and no input. The model is not lying. It is obeying.
The failure mode I keep hitting is the silent 200. An API returns HTTP 200 with a well-formed JSON body and an empty content field. Upstream is healthy by every metric a load balancer understands. Downstream, three hops away, an analyst reads a confident paragraph about a project that had no news that day.
Verification cost is not a theoretical concern. During the 2022 drawdown I spent six months optimizing zk-SNARK circuits for a mid-sized rollup, cutting proof generation time by 15%. Fifteen percent sounds trivial until you price it under panic conditions, when the marginal user leaving a chain is leaving because the exit felt slow. Verification is a stabilizer. So is refusal.
I learned this the expensive way in 2017, auditing ERC-20 contracts for my undergraduate thesis โ fifty-plus ICOs, forty hours a week, three critical reentrancy vulnerabilities. The finding that changed how I work was not the bugs. It was the number of contracts that returned true from a function that had done nothing. Silence, in Solidity, is a return value. I have treated nulls as evidence ever since. Where code becomes law in the digital frontier, the empty case is the case that gets you.
The same pathology runs through payments data. Stablecoin corridor volume in emerging markets is routinely self-reported by the institutions that profit from the number looking large; the chain is a settlement record, not a demand record. My work on corridor flows keeps landing in the same place โ the users are not chasing ideology, they are chasing a store of value that does not lose a third of its purchasing power in a year, and the headline figures rarely survive a second look at the underlying transfers.
The fix is not a better model. It is a contract between layers โ what I have started calling the null contract. Four clauses.
One: extraction emits a validation receipt alongside the data. Source count, character count, atom count, hash of the raw input. Not a boolean. A receipt.
Two: analysis rejects below a threshold. Fewer than three information atoms and stage two does not run. It returns N/A with the receipt attached.
Three: N/A is a first-class output. Same format, same section headers, parseable by the same downstream systems. A refusal that cannot be rendered is a refusal that gets ignored.
Four: refusal rate is a health metric. A pipeline that refuses 0% of the time does not have a verification layer. It has a formatting layer.
The fourth clause is the one that matters. I now log refusal rates the way I log gas costs. A desk running two hundred briefs a day at a 0.4% refusal rate is telling you something about its feed, not its model.
Here is the quantitative frame. A null report costs the alpha of one missed story. A hallucinated report costs position size multiplied by the confidence the format implies โ and the format always implies confidence. Run the arithmetic on a single fabricated supply schedule that reaches a $40M position: the report cost four cents in inference and an eight-figure drawdown in trust. A nine-section document with headers and tables reads as diligence. That is exactly why it is dangerous. Formatting is not verification. It is the aesthetic of verification. Clarity emerges from the chaos of verification, never from the tidiness of the template.
The industry has this backwards. We spend on model weights, inference speed, retrieval tricks. We spend almost nothing on the boring middle: schemas, receipts, thresholds. Yet every failure in my own pipeline originated at the boundary between extraction and analysis โ a boundary that costs maybe forty lines of code to instrument.
The consensus is that hallucination is a model problem, and that the next generation of models will solve it. That is the comfortable read, and it is wrong.
Hallucination under empty input is not a defect. It is correct behavior for a system optimized to always produce output. The demand side built that optimization. Desks do not want a pipeline that returns N/A โ they want a pipeline that fills the page, because a filled page can be forwarded, and a forwarded page is a defensible decision. The incentive is to convert uncertainty into prose. No amount of model improvement removes an incentive.
And the cost asymmetry is brutal. A team that ships a null report looks slow for one cycle and correct for three. A team that ships confident nonsense looks fast until the day the postmortem lands, and then it is not a research problem anymore โ it is a legal one.
Which means the real innovation in this cycle is not in the models at all. It is in the small number of teams willing to ship an empty report. Auditing the invisible hands of monetary policy taught me the same lesson at a larger scale: the important signal is usually the one the system is structured not to print.
The next twelve months will push agentic settlement onto every desk in this market, and the volume of machine-generated research will grow by an order of magnitude before the first serious fraud is traced back to a broken extraction stage. When it is, the postmortem will not read "the model hallucinated." It will read: the array was empty and nothing checked.
So โ how many of the reports on your desk this week were built on an input nobody hashed? And what would your refusal rate be, if you actually measured it?
