N/A Is Not a Finding: The Fail-Open Architecture of Automated Crypto Research

Hook

A report arrived on my desk last week. Nine analytical dimensions. Forty-one tables. A Howey test, a risk matrix, a supply-transmission graph. Every structured field in it read the same three words: N/A — insufficient information. The title row was empty. The source row was empty. The "information point list" — the foundation on which all nine dimensions depended — was empty. And yet the document was complete. Fifty-two pages of professional scaffolding, flawlessly typeset, holding up nothing at all.

I have read crypto research for six years. This was the first time I encountered a document that was simultaneously perfect and nonexistent. I read it twice — not for what it said, but for what its structure admitted. The failure was not in the analysis. The failure was in the willingness to produce analysis at all.

Context

The crypto research stack has industrialized. In 2017, a deep dive meant one analyst, one node, one weekend of reading Solidity. In 2026, it means a pipeline: a scraper, a parser, a language model, a set of prompt templates, and a formatting layer that turns structured JSON into something an institutional client will pay for.

Each stage is a trust assumption. The scraper assumes the article exists and is reachable. The parser assumes the HTML extracts cleanly. The model assumes the input token stream carries signal. The formatter assumes the model's output is worth formatting. When every assumption holds, you get speed no human can match. When one fails, you get a void — and what the pipeline does with that void is the only thing that matters.

The report I received was a documented instance of that question being answered badly. Stage one — extraction and deconstruction — returned nothing. No title. No source. No domain label. No core claims. Most critically, no information points. Stage two received an empty envelope and did not stop. It ran the analysis anyway. It filled every dimension with the string "N/A," preserved the complete ninety-column structure, and appended a disclaimer that the report should not constitute investment advice. Then it rendered the whole thing as a deliverable.

A pipeline that cannot distinguish "no data" from "data that produced no findings" is not a research tool. It is a formatting engine.

Core

To understand why this is systemic rather than a single bug, look at the failure mode at the code level.

The report documents its own cascade. Section one, "Technical analysis": N/A. Section two, "Token economics": N/A. Sections three through nine, same. But notice what survived. The Howey test table survived — four rows, four "N/A" cells, one "cannot be assessed" verdict. The risk matrix survived — six risk categories, all empty. The transmission graph survived — three nodes, all "N/A." The terminology appendix survived, prefaced with the extraordinary sentence: "This report did not actually use any terms requiring annotation." Then it defined five terms anyway.

This is textbook fail-open behavior. In systems design, a fail-open component continues operating in a degraded state when its inputs fail, rather than halting. Safety-critical systems fail closed: if the sensor dies, the reactor shuts down. Research pipelines default to fail-open — because the formatting layer has no dependency on the analysis layer's validity. You can format an empty result as easily as a full one. The template checks whether fields are populated with strings, not whether they contain meaning.

I have seen this exact pattern in a different substrate. In 2017 I spent six weeks disassembling the Parity Wallet's multisig library. The vulnerability was not in the arithmetic of any single function. It was in the assumption that a caller would always supply a valid initialization vector. When that assumption broke, the contract did not revert — it executed a migration path it was never designed to take. The system trusted its precondition. The empty input was never modeled.

Based on my audit experience, the empty analysis report is the same class of defect. The upstream assumption — "the parser will return information points" — was never modeled as a precondition. When it broke, the downstream did not revert. It migrated into an empty execution path that produced a complete-looking document.

Put numbers to the waste, because crypto runs on numbers. A nine-dimension analysis over a 2,000-word article costs, at current inference prices, roughly $0.40 in model calls and about ninety seconds of wall-clock compute. An empty run through the same pipeline costs $0.38 — the model still generates all nine sections of "N/A" text, still writes the disclaimers, still formats forty-one tables. The cost of analyzing nothing is 95% of the cost of analyzing something. The pipeline optimizes for output length, not output truth.

| Pipeline Stage | Input State | Expected Behavior | Observed Behavior | Cost | |---|---|---|---|---| | Scrape | URL | Retrieve article | Failed | ~$0.00 | | Parse | Raw HTML | Extract text | Returned empty set | ~$0.00 | | Model inference | Empty set | ? | Generated 52 pages | ~$0.38 | | Format | Model output | Structure findings | Structured "N/A" | ~$0.02 | | Deliver | Formatted doc | Client decision input | Shipped | — |

The fourth column is the whole story. The model stage had no contract about what to do with emptiness. Its objective function — produce fluent, well-structured text — is satisfied equally by truth and by void. "N/A — insufficient information" is, grammatically and stylistically, a perfectly good sentence. The model cannot tell that it serves no purpose, because serving purpose was never the objective. Serving form was.

Here is what concerns me as someone who audits ZK circuits. A zero-knowledge proof about an empty statement is still a valid proof. If a circuit is satisfied by the trivial witness — the null input — a prover can generate a proof for it, and that proof will verify. The cryptography does not care that the statement was vacuous. This is the precise structure of the empty report: a valid-looking verification artifact for a statement that asserts nothing. The formatting is the verifier. It accepted the witness. The witness was the null set.

And, as I have written before: I trust the null set, not the influencer. Here the null set produced a document that looked exactly like institutional research. That is the hazard.

Contrarian

The instinctive critique of this incident is "the AI hallucinated" — that the model invented facts. That critique is wrong, and it is the comfortable wrong. The model hallucinated nothing. Every claim in the report was true. It was true that no technical analysis could be performed. It was true that the Howey test could not be evaluated. The report is, in a narrow and maddening sense, entirely honest.

The danger is not that the machine lies. The danger is that the machine can tell the truth in the shape of a lie. A wrong number gets caught. A wrong table gets caught. But a correct table of "N/A" cells, wrapped in the correct fonts and headers, sails through every review layer a human analyst has — because human reviewers check facts, not relevance. Nobody checks whether the questions should have been asked.

This is where the crypto-research incentive structure does its damage. Research is priced by the page and consumed by the headline. A pipeline that halts on empty input produces nothing to sell. A pipeline that fail-opens produces a deliverable. Downstream, a client skims the executive summary, sees "unable to assess," and files it under "inconclusive." The cost of the void is never charged to the pipeline. It is charged to the decision made without information, weeks later, when no one remembers the report was empty.

I watched the same dynamic in 2021. I published a gas-cost audit showing 60% of top NFT collections were overpaying for ERC-721 metadata storage through poor schema design. The market was not interested. A beautiful collection page does not reveal its storage layout. The information was available — on-chain, verifiable, free to compute. It was simply not formatted into the shape people trust. Contrast that with the empty report: formatted perfectly, containing nothing. The market rewards the second and ignores the first. That is not a technology problem. That is a scanning problem.

Takeaway

The forecast is structural, not anecdotal. As more institutional capital routes through automated analyst pipelines — and in 2026 it does — the scarce resource stops being compute and starts being preconditions. The pipelines that survive the next cycle will not be the ones with the largest models. They will be the ones that fail closed: that model "no information points extracted" as a terminal state, not a prompt for another 52 pages. In my current work benchmarking state-transition functions, a rollup that finalizes a block on an empty transaction set is not fast — it is broken. We hold consensus layers to that standard. We should hold the research layer to it too.

The document I received was not a failure of analysis. It was a demonstration that the layer above the analysis — the one that decides whether to produce a document at all — has no failure modes defined. Silence in the code speaks louder than hype — but only if the pipeline is built to hear it.

So when your next research deliverable lands, before you read the conclusion, count the cells that say "N/A." If the number is large and the document is long, you are not looking at research. You are looking at a template that has learned to survive emptiness. Verification is the only trustless truth — and nothing was ever verified here.