Null Input, Full Confidence: The Provenance Gap in AI Crypto Research

Last week I fed a research agent an empty brief and it handed me 1,400 words back.

Not empty as in thin. Empty as in zero. The title field was blank. The source field was blank. The list of extracted information points — the thing every downstream claim is supposed to hang off — was an empty array. The thesis field was null.

The agent returned a nine-dimension analysis anyway. Every dimension had a conclusion, an evidence line, a hidden-information note, a risk flag, and a confidence score somewhere between 0.72 and 0.85.

One field told the truth. Domain tag: unclassified.

That's the entire story compressed into a single default value. The upstream classifier never ran. Everything downstream of it was prior — genre-shaped, plausible, unfalsifiable. I didn't need a second read to know what I was looking at, because I've written the same class of bug myself and paid for it in P&L.

This isn't a story about one broken script. It's about the default behavior of nearly every LLM-based research tool that shipped into this cycle, and about why the bull market hasn't noticed yet.

The Framework Was Never the Problem

The nine-dimension discipline is genuinely good. I want to be clear about that before I take it apart.

Every conclusion has to name its basis. Every basis has to point to a numbered information point. Every dimension carries a confidence value and a risk flag. That four-part structure — conclusion, evidence, hidden information, risk — forces an analyst to separate what they observed from what they inferred. Most crypto research does not do this. Most crypto research is a thesis with a chart taped to it.

So the schema is not the failure. The schema is the thing that makes the failure visible.

Here's the hard dependency nobody budgets for: provenance. Nine dimensions, four cells each, is thirty-six cells. Thirty-six cells want filling. A human analyst staring at an empty brief writes "insufficient data" in the first box, sighs, and goes to get coffee. A language model fills all thirty-six, because that is what it was trained to do and there is no cell in the schema labeled "refuse."

The timing matters. Through 2025 and into 2026, every desk, every fund, every newsletter shipped an agent. Structure sells, because structure photographs well. A dashboard with nine panels and a confidence column looks like work. In a bull market nobody audits the panels — they screenshot them and forward them to the group chat.

I came to this from the mempool, not from the research desk. In August 2020 I ran a Python bot against Uniswap V2 swaps, watching pending transactions, bidding gas against everyone else doing the same thing. It executed 140 transactions in a single block at peak and cleared $85,000 in three days. It also nearly got my IP blacklisted by two RPC providers because I'd tuned the gas bidding so aggressively that I was degrading my own connectivity. The lesson I took out of that wasn't about strategy. It was that my logic was fine and my inputs were garbage — stale nonces, cached state, providers serving me a view of the chain that was four hundred milliseconds old. I was trading on ghosts and winning, which is the most dangerous possible feedback loop.

AI research agents have the same disease, one layer up.

The Conditional Distribution Is Peaked on Genre

When you hand a model a schema with empty inputs, you are not sampling uniformly from the space of possible analyses. You are sampling from the space of analyses that look like this. The training distribution is dense with exactly this artifact — confident, structured, hedged in the right places, formatted to spec. The model is not lying to you. It is continuing a document. Next-token prediction under a helpfulness prior does not have a stop condition called "the evidence is absent."

A schema with empty inputs is a request for fiction, phrased politely.

And the fiction is high quality. That's the part that should worry you. It reads like the median analyst's output because it was trained on the median analyst's output. The evidence lines restate the conclusions in slightly different words. The hidden-information sections contain no new facts, only implications of facts that were never established. The risk flags flag risks that are always true — regulatory uncertainty, smart contract exposure, liquidity fragmentation — which is to say, risks that carry zero information because they apply to every asset in the sector.

You can spot it if you know what to look for.

Null Propagation Has a Signature

In any pipeline, missing data degrades to a default, and that default survives all the way to the surface. Learning to read the defaults is most of the skill.

"Unclassified" is the default. So is a confidence band that clusters tightly — 0.72 to 0.85 is not a distribution over evidence, it is a distribution over tone. Real evidence produces dispersion: 0.94 where you have a signed transaction and an on-chain receipt, 0.31 where you have a rumor. When everything lands in the same narrow band, the number isn't measuring confidence. It's measuring how the model has learned to sound.

So is an evidence line that paraphrases its own conclusion. So is a "hidden information" block with no discrete fact in it. So is a domain tag that resolves to a placeholder.

The engineering fix is boring and almost nobody ships it: assert at ingest. If a required field is empty, halt the pipeline. Do not degrade gracefully. Graceful degradation is how an empty array becomes a nine-dimension report.