Fourteen pages arrived in my inbox on a Tuesday morning. Nine analytical dimensions, each with its own heading, each with a status field, and every status field reading the same three words: insufficient information. The model had not failed to produce a document. It had failed to find a fact โ and then it said so, at length, in a table, with headers. Eleven of the fourteen briefs in that pilot came back the same way: structurally immaculate, semantically hollow. Half the room wanted to debug the model. I wanted to sit with the eleven blank fields, because a system that can render its own ignorance in perfect formatting has just told you something uncomfortable about every dashboard you trust.
The pipeline was ordinary. Stage one reads a source and extracts information points โ names, parameters, treasury figures, unlock schedules, the granular stuff. Stage two takes that list and reasons across nine dimensions: technical, tokenomic, market, ecosystem position, regulatory, team and governance, risk, narrative, and supply-chain transmission. Stage one returned an empty list. Stage two, to its credit, declined to invent. It filled every cell with N/A and flagged the input pipe as defective.
That refusal is the entire story, and it is rarer than you think. In most production deployments, the abstention never reaches a human being. It gets smoothed. A model fine-tuned on engagement will happily write nine confident paragraphs from zero facts, because the loss function rewards fluency and nobody scores it on whether the treasury number was real.
I have spent twenty-five years watching this pattern in slower motion. In late 2017 I audited more than fifty whitepapers for legitimacy rather than code, and published a comparative piece on three ICOs that promised decentralization while keeping treasury control in a single unlabeled wallet. Fifteen thousand people read it within a week. The lesson was not that the teams were dishonest. It was that technical brilliance without ethical governance collapses, and that collapse is always preceded by a document that looked complete.
The empty table matters now because of where these documents land. Through 2024 and 2025, DAOs bought summarization tooling to survive proposal volume. Contributor budgets were cut in the drawdown, so human reviewers per proposal fell from three to one to a volunteer with a day job. Automation rose exactly as scrutiny fell. That is not a coincidence; it is the procurement logic. It is the same logic that put AI agents into delegate seats, and the accountability standards we argued for in 2026 were drafted faster than they were adopted.
In 2020 I ran twelve workshops teaching non-technical users how to read Aave's risk parameters. What lost people was never complexity. It was false confidence. A confident sentence with a wrong number moved more people into bad positions than any wall of jargon. The 2022 newsletter taught me the same lesson from the other side: panic is not caused by bad news, it is caused by discovering that what you were shown was not what was there.
Here is the mechanism I want you to see clearly. A schema-valid document is not a semantically valid one, and crypto's tooling has spent a decade confusing the two. JSON is a promise about shape. It says nothing about truth. When stage one returns nothing and stage two returns nine dimensions of N/A with confidence scores attached, the output passes every structural check in the stack and fails the only check that matters. Worse, the N/A cells read as rigor. They look like an analyst who considered each angle and found the evidence wanting.
I watched a version of this in an L2 governance cycle last quarter. A sequencer key rotation came up for a vote โ a five-of-nine multisig reconfiguration, the kind of proposal where the real question is who can halt the chain and under what conditions. The agent summary circulated to token holders said no material risk identified. It said that because the risk extraction field had come back empty. Nobody lied. The confidence score was 0.9. Seven hundred wallets voted on the strength of a paragraph generated from a blank input. Decentralized sequencing has been a slide deck for two years; the operators are still a handful of people with keys, and now the summaries describing them are generated from nothing.
The same failure mode is spreading into places with less tolerance for it. Post-ETF Bitcoin flows moved the interesting governance questions into custody rails, where the people asking them are compliance officers rather than communities. Compliance tooling is precisely where a beautifully formatted report with empty fields gets signed and filed. Peer-to-peer electronic cash became a settlement asset in a wrapper, and the paperwork around the wrapper is what decides who gets access.
Underneath all of it sits the oldest problem in this industry. Code is law until an upgrade needs to ship, and then the upgrade right sits with a few multisig admins accountable to nobody in particular. Every summary that says no material risk identified is a summary of a system whose actual authority is a quorum of people you cannot name.
So what does good look like? The absence of data has to become a first-class, signed event. Not a blank cell. An attestation: this dimension was queried, the source returned nothing, here is the hash of the query, here is the timestamp. If a governance agent votes on a proposal it could not parse, that abstention should be on-chain, attributable, and scored. I have been pushing this in the three DAOs I work with since the interface protocol work in 2024, and the resistance is always the same โ it makes the dashboards look worse. Good. The dashboards should look worse. Trust is earned in bear markets, and right now the bear market is the only thing keeping anyone honest.
Now the contrarian part, because I do not want to be read as celebrating abstention.
A model that says I don't know is more trustworthy than one that improvises, but only if the abstention is calibrated. An agent that abstains on everything is useless and, worse, is abdicating. The interesting metric is not abstention rate. It is the correlation between abstention and actual information scarcity. If the system abstains when the data is thin and commits when the data is thick, it has learned something. If it abstains whenever a document is longer than its context window, it has learned to hide. Almost nobody measures this, because the metric requires ground truth about what was knowable, and building that ground truth is expensive, unglamorous work that no one funds in a drawdown.
The deeper contrarian point is about the template itself. Nine dimensions is coverage theater. A blank page frightens people into asking questions. A filled grid โ every cell present, every cell marked N/A, every row accounted for โ reassures them into silence. The format did the lying, not the model. We built the nine-box grid because it looked like diligence, and now the grid is the product and the content is optional. Empathy is the ultimate security layer, and the most empathetic thing a governance tool can do is refuse to hand a tired volunteer a document that looks finished when it isn't.
What I am watching from here is narrow and concrete. Whether the next generation of DAO tooling ships abstention attestations as a primitive rather than an error state. Whether delegate agents get scored on calibration instead of output volume. Whether any L2 publishes the human-readable rationale for its sequencer key changes, with the actual signers named, before the vote rather than after the incident. None of that is a bull case. It is a survival case, and in a market where treasuries are bleeding and participation is thin, survival is the only thesis that pays.
People first, protocol second. Always. The fourteen-page document that said nothing was the most honest thing I read all week. The question is whether your governance stack would have let it reach you.