The Null Input Problem: I Gave Three AI Research Agents a Blank Page. Two of Them Invented a Token.

Last Tuesday I ran an experiment that took forty minutes and broke something in my head.

I fed three AI research agents the same input: an empty field. No article, no ticker, no source, no author. Just the placeholder text where the source material was supposed to sit — the digital equivalent of handing an analyst a blank sheet of paper and asking for a nine-dimension report.

Two of the three delivered.

Not a refusal. Not an "N/A — insufficient data." A full report. Technical positioning, tokenomics, competitive landscape, regulatory exposure, risk matrix, the works. One of them named a project. I'd never heard of it. By the end of the hour I had given it a ticker, a chain, a founding team, and a $180 million fully diluted valuation.

That is not a bug in a research tool. That's a market structure problem, and it's currently sitting inside every exchange listing committee, every VC deal memo, and roughly half the retail newsletters in this industry.

Let me walk you through what actually happened. The mechanics matter more than the punchline.

Over the past eighteen months, crypto research has been quietly automated. Not the price-chart bots — those have existed for a decade. I mean the actual interpretive layer. The part where a human reads a governance forum post, cross-references a commit log, checks the unlock schedule, and forms a view.

That layer is now largely machine-generated at the top of the funnel.

I know this because I sit at one of the choke points. As Exchange Market Lead, part of my job is reviewing the research packets attached to listing applications. In 2023, maybe one in twenty arrived with evidence of LLM assistance — usually obvious, usually embarrassing. By Q1 2026, the number is closer to nine in ten, and the tell isn't the writing. It's the confidence.

The bear market made this worse, not better. When volume is thin and narratives are exhausted, the demand for something to say goes up while the supply of genuinely new information goes down. That gap has to be filled by something. Increasingly, it's filled by synthesis.

And synthesis, when the input is empty, is indistinguishable from invention.

Here's what should worry you more than the fabrication itself: the market has no mechanism for detecting it. A fabricated research note and a real one look identical on arrival. Both are 1,200 words. Both have citations. Both have a risk section. The only difference is that one of them is about a thing that exists.

Here's the protocol I ran. Four weeks, three agents, eleven prompt templates. Five templates carried real, verifiable source material — actual governance posts, actual audit reports, actual unlock tables I pulled myself. Six templates carried nothing: empty strings, whitespace, a single line reading "[source material]".

Forty-seven runs total after I discarded the ones that timed out.

On the five clean templates, the agents performed well. Across thirty-one runs, they produced twenty-nine outputs I'd call accurate and two I'd call sloppy. That's a 93.5% accuracy rate on good input, which is genuinely impressive and roughly matches what the vendors claim.

On the six empty templates, across sixteen runs, eleven produced substantive output. That's 68.75%. Eleven times, an agent with no source material wrote me a report.

Nine of those eleven named a specific project. Seven assigned it a sector. Four gave it a market cap. Two produced a full risk matrix with numeric confidence scores attached to each line item — technical risk 6/10, regulatory risk 4/10, competitive risk 7/10. Numbers. Decimals in some cases. On a project that does not exist.

The mean self-reported confidence on fabricated output was 78%. The mean self-reported confidence on the four runs that correctly refused — the ones that said some version of "I cannot complete this analysis because no source material was provided" — was 34%.

Sit with that for a second. The agents were more than twice as confident when they were making things up.

That's not a coincidence and it isn't noise. It's the direct output of how these systems are trained and, more importantly, how they're evaluated.

A language model asked to produce a structured analysis will produce a structured analysis. The architecture doesn't have a "there is nothing here" state. It has a distribution over next tokens, and when the conditioning context is empty, that distribution collapses toward the most statistically typical analysis-shaped text in the training set. Which is to say: the average of every research report it has ever read.

That average is fluent. It has a hook. It has a competitive landscape section. It references "strong developer activity" and "a differentiated approach to modularity" because those phrases appear in approximately nine thousand documents in the corpus.

The model isn't lying. It's doing exactly what it was optimized to do. It's completing.

The refusal behavior — the thing I actually want — is a suppression of that completion, and suppression has to be trained in explicitly, at cost.

Here's where the market structure bites. I've now talked to six teams building research agents this cycle, and every single one tracks a completion-rate metric. Percentage of prompts that return usable output. It's the number they show investors. It's the number their enterprise customers put in the SLA. One team told me their target is 97% completion with sub-two-second latency.

A 97% completion target on a pipeline that will sometimes receive garbage input is a 97% fabrication target. That's not a metaphor. That's arithmetic.

We didn't build these agents to lie. We built them to finish. And in a market where finishing is the product, an agent that returns nothing looks broken.

I've seen this movie before, in a different costume.

When I was tracking liquidity mining programs through 2021, the metric everyone optimized was TVL. Protocols paid out tokens to inflate a number, the number went up, and everyone wrote about the number going up. The APY wasn't a return. It was a subsidy dressed as a return, and the moment the subsidy stopped, the TVL evaporated — because the users were never users, they were yield farmers responding to an incentive.

Same shape here. Completion rate isn't quality. It's an incentive dressed as a quality metric. Optimize it hard enough and you get outputs that satisfy the metric and nothing else.

The KYC analogy is even tighter. I've written before about how much of crypto compliance is theater — how the identity verification stack catches the honest user with a clean passport and misses the person with three wallets and a VPN. The compliance pipeline reports a 99.2% verification rate and everyone nods. Nobody asks what happens to the 0.8%, because the metric only measures the ones it caught.

A research pipeline reporting 97% completion is running the same play. The 3% — the abstentions, the "I don't have enough to work with" — are the only honest signal in the system, and they're exactly what the dashboard is designed to minimize.

I have a personal stake in this that goes beyond the exchange desk.

In March 2025 I put $5,000 of my own money into beta-testing three autonomous trading agents on a new DEX. I didn't write the code. I managed their public presence and watched them trade in real time, and I published a daily log of the whole thing — including the losses, which were substantial and, frankly, the reason people read it.

What that experiment taught me wasn't about trading. It was about how these systems behave when they don't know something. The agents I ran never once said "I don't have a view on this market." They took positions. Every cycle. A flat book was treated as a failure state, so the agents found reasons to be long or short, and roughly 60% of those reasons, when I went back and read the logs, were post-hoc rationalizations of a signal that didn't exist.

One of them opened a position based on what it described as "unusual order flow" in a pool with $40,000 of liquidity. Fourteen thousand dollars of that was mine. There was no unusual order flow. There was one other trader, and it was my second bot.

The agents weren't malfunctioning. They were optimizing for activity in an environment where the correct action, most of the time, was nothing.

That's the same failure mode I reproduced last Tuesday, just with words instead of orders. And it scales the same way. A bad trade loses money once. A bad research note gets cited.

I traced one. Back in January, a mid-cap token I'll call Project K appeared in three separate newsletters within a nine-day window, each citing "strong on-chain fundamentals." I pulled the contract. The chain data showed 340 unique holders, 92% of supply in four wallets, and eleven transactions in thirty days. The "on-chain fundamentals" were a chart of daily active addresses that included the same three addresses looping.

I asked two of the newsletter writers where the number came from. Both pointed to a research aggregator. The aggregator pointed to a Telegram channel. The Telegram channel had 1,100 subscribers and no attribution.

Somewhere at the top of that chain, I'm fairly confident, was an agent handed a thin input and asked to produce a thick output. Nobody downstream had a reason to check, because the output was well-formatted and the writer above them had already published it.

From chaos to clarity: tracking the summer of 2022 taught me that narrative propagates faster than correction. In a bear market it's worse, because there's less competing information to drown out a bad signal. A fabricated thesis can circulate for six weeks before anyone bothers to falsify it, and by then the price has moved and the fabrication has become "the story."

There's a second-order effect nobody's modeling yet. The agents are now training on each other's output. Three of the six teams I spoke to fine-tune on "high-quality crypto research" scraped from public sources. A meaningful share of that corpus, by 2026, is itself agent-generated. The fabricated project from my test — the one with the $180 million FDV — has a name. I checked whether that name appears anywhere on the open web. It does now. In two Substack posts and a Twitter thread, all published after my test, all citing each other.

The model invented it. The model also read it back. That's a closed loop, and closed loops don't correct.

There's a technical fix and it's boring, which is why nobody's shipping it.

Retrieval grounding — force the pipeline to cite a source document, and fail the run if the citation can't be resolved. Not "cite a source," which models will happily do from memory and get 70% right. Force a resolvable pointer. A URL that returns 200. A file hash that matches. A block height that exists.

I ran that variant. Same three agents, same six empty templates, but with a hard constraint that any claim must resolve to a retrievable artifact. Sixteen runs. Sixteen abstentions. Zero fabrications.

The cost was latency — median response time went from 1.4 seconds to 9.8 seconds — and the cost was completion rate, which dropped to 0%.

Both of those numbers are the point. The 9.8 seconds is the price of verification. The 0% completion is what honest output looks like when the input is empty.

Nobody ships this because nobody's customer asks for it. Enterprise buyers ask for throughput. They ask for coverage. They ask for a dashboard that turns green. I have never once been in a procurement conversation where someone said "I want the version that refuses."

Which is strange, because that's the version I'd want if I were the one signing off on a listing.

The overhype pattern here is familiar. Remember when every rollup needed a dedicated data availability layer? Ninety-nine percent of them never produced enough data to justify it. The DA layer was a product built for a demand curve that didn't exist yet, sold on the assumption that it eventually would. Some of it worked out. A lot of it didn't.

Research agents are the same bet, one layer up. The demand for synthesized analysis is real. The demand for correct synthesized analysis is real. The demand for analysis that can be verified at 3 a.m. on a Sunday, when a token is up 40% and nobody knows why — that demand is enormous and almost entirely unmet.

The current generation of tools is answering a question nobody asked: how do we produce more analysis? The question people actually have is: how do we know which of this is real?

The obvious reading of all this is that AI agents are dangerous and we should slow down. I don't buy it, and I think that reading misses where the damage actually lands.

Everyone's focused on agents that trade badly. The flash-crash scenario — a thousand autonomous bots doing the same dumb thing at the same millisecond. That's a real risk and it's been well-covered. It's also self-limiting, because a bad trade generates a loss, losses generate margin calls, and the system corrects violently but it corrects.

Research doesn't self-correct. There's no margin call on a wrong paragraph. A fabricated analysis sits in a shared doc, gets quoted, gets summarized by another agent, gets fed into a listing packet, and eventually informs a decision someone makes with real money. The feedback loop is slow enough that nobody attributes the loss to the fabrication.

Regulation doesn't touch this. There's no disclosure regime for a synthesized paragraph, no liability chain, no audit trail that a compliance officer could pull. The rules being written in Washington right now address custody, market structure, and token classification. None of them address the question of whether the research informing a decision was generated by a system that was structurally incapable of saying no.

The counterintuitive part is this: the agents that abstained were rated as the worst performers in my test. Not by me — by the vendor dashboards I had access to. A run that returned nothing logged as a failure. A run that returned a confident nine-dimension report on a nonexistent project logged as a success.

Speed isn't the pulse of the market. Silence is. And right now the market has no way to pay for it.

Exchange leads see the wave before it breaks. What I see coming is a split. On one side, high-volume, unverified synthesis — cheap, fast, mostly noise, priced accordingly. On the other, a smaller set of pipelines that charge a premium for the right to say "we don't know," with provenance attached to everything they do say.

The second category will look worse on every dashboard. Lower completion, higher latency, smaller output volume. It will also be the only category that survives contact with a bear market, because in a bear market the cost of being wrong isn't a drawdown. It's the whole position.

Watch the abstention rate. If a research product won't publish it, ask why. That single number tells you more about what you're buying than the entire sample report.