Last week a research pipeline I helped design returned a nine-row matrix in which every cell carried the same three words: insufficient information. Technical surface, token economics, market structure, ecosystem position, regulatory exposure, team governance, risk profile, narrative expectation, supply-chain dependencies — nine dimensions, nine blanks. The document it had been handed contained no project name, no source, no dated claim, no figure. And instead of producing the 1,500-word analysis it had been asked for, it produced a refusal, formatted as a table, with a note explaining that any conclusion drawn from that input would be invention dressed as diligence.
I have read several thousand machine-generated crypto reports in the past eighteen months. This was the first one I trusted on sight.
That sounds like a joke at my own industry's expense. It is not. In a market where every dashboard is green-adjacent and every thread terminates in a price target, the most informative artifact I encountered this quarter was a document that said nothing — and said so with precision.
The research economy did not collapse in the bear market. It industrialised. When my publication's revenue fell 70% in 2022, the pressure I felt was not to write better; it was to write more. The teams that survived learned to publish on volume, and by 2025 much of that volume was being produced by models. Retrieval-augmented pipelines now ingest a press release, a governance forum thread, an on-chain dashboard and a price feed, then emit a structured deep analysis in eleven seconds. The format is familiar because it is uniform: nine dimensions, a summary paragraph, a risk section that lists risks nobody has ever actually been exposed to.
What those systems optimise is completion. A model trained to be helpful has been taught, at enormous expense, that an unanswered question is a failure. So when the retrieval layer comes back thin — a document with no dates, a token with no disclosed supply schedule, a team with no verifiable identity — the model does not report the gap. It fills it. It writes that the team has a strong technical background because the sentence carries high probability mass, not because anyone checked.
I spent three weeks in 2022 auditing the Terra collapse for a post-mortem I called Narrative Decay, and the finding that outlived the report was not about code. It was that broken promises erode trust faster than broken code does. A model that fills gaps is a promise-breaking machine running at industrial speed. The nine-row blank table is what it looks like when one of them is finally built to break the habit.
Abstention is not a personality trait a model has. It is a decision boundary, and boundaries have to be placed deliberately. In the pipelines I have reviewed, the boundary is almost never set by the model's confidence. It is set by the product requirement: an empty cell does not get clicked, does not get quoted, does not get syndicated. The economic incentive runs entirely one way — toward the filled cell — which means abstention must be engineered against the market, not with it.
Here is the asymmetry that explains the whole category. A fabricated analysis imposes its cost on the reader: the reader allocates capital, the reader misjudges a counterparty, the reader holds a position through a drawdown they never understood. An abstention imposes its cost on the publisher: the publisher ships a document that looks thin next to a competitor's confident nine-paragraph take. Cost is displaced from the party that produces the error to the party that consumes it. Markets do not price that externality. They over-supply the error.
I first understood this in 2017, sitting with seventeen ICO whitepapers and a cybersecurity degree. Six months of reading produced three projects carrying smart contract vulnerabilities that were later exploited in exactly the way the code suggested. But the part of that work I still defend is the appendix: a list of claims I could not verify — no repository, no vesting schedule, no identifiable engineer — published as unknowns rather than resolved by inference. Readers hated the appendix. It was the only section that aged well.
The same failure appears on-chain, and it is measurable.
Take total value locked, the industry's favourite number. It counts the same dollar twice when it is deposited through a liquid staking receipt and again as collateral. It counts a token whose price collapsed as though quantity were value. It rises when a protocol pays incentives to rent deposits that leave the week emissions stop. None of this is fraud; all of it is a filled cell where the honest entry was "unknown, and the definition matters."
Take layer-2 chain counts, which the market reads as evidence of technical superiority. I have argued for two years that the real difference between OP Stack and ZK Stack is not the proof system — it is who can persuade more projects to deploy a chain first. That is a distribution contest wearing a cryptography costume. The metric measures salesmanship and reports it as engineering.
Take Bitcoin's inscription and Runes activity, celebrated as network utility. BRC-20 and Runes are a Rolls-Royce hauling cargo: the car is insulted and the freight is small. The throughput is real; the claim that it constitutes demand for Bitcoin's monetary function is a filled cell with nothing behind it.
And take the regulatory narrative I have watched most closely. Hong Kong's virtual asset licensing regime is routinely described as an embrace of innovation. Read the licensing thresholds, the custody requirements, the token admission rules, and the timing relative to Singapore's framework, and a narrower reading holds: this is jurisdictional competition for the Asian financial centre's flows, conducted through compliance architecture. That reading is defensible. It is simply not the reading the press release supplies.
Every one of those cases has the same shape. The data exists. The definition is contested. The pipeline resolves the contest silently and prints a number.
What would a calibrated system do instead? It would attach the retrieval trace to the claim — not a citation, a trace: which document, which timestamp, which field was absent. Provenance is not a footnote; it is the difference between a measurement and an assertion. In 2026 I helped build a pilot that authenticated a thousand articles from independent journalists using zero-knowledge proofs of authorship. The lesson that transferred was not about cryptography. It was that verification has to be attached at the point of creation, because it cannot be reconstructed afterwards. An analysis without a trace is a rumour with a chart.
Here is where I stop defending the empty table.
Abstention has a shadow. A system that says insufficient information can be a system that has not looked. The nine-row matrix was honest, but honesty was cheap that day — the input was empty, and refusing to analyse nothing costs nothing. The harder test is a document with a name, a token and a promise, where the retrieval layer must do the labour of finding the filing, the wallet, the engineer. Refusing to fabricate is half of integrity. The other half is the work of finding out, and the first half is frequently used to avoid the second.
I have watched insufficient data become a shield in newsrooms and audit reports alike. It protects the analyst's reputation while leaving the reader exactly as uninformed as before. In a bear market, where the reader's real question is whether their assets are safe, silence is not neutrality. Silence is a position, and it is usually the position of whoever benefits from the status quo.
Soulless finance is just empty pixels. But so is a blank table with no one standing behind it.
The next narrative in this industry will not be about which chain is faster or which proof is smaller. It will be about the provenance of the claim — who measured it, from what, and what they declined to assert. Code doesn't hesitate; it executes whatever it was told, including the instruction to invent. Trust must be engineered, not promised. The question I keep returning to is not whether our tools can learn to say I don't know. It is whether we can build a market that pays them for it.